Abstract
The simulation of complex engineering applications often requires the consideration of componentlevel dynamics whose nature and timescale differ across the elements of which the system is composed. Cosimulation offers an effective approach to deal with the modelling and numerical integration of such assemblies by assigning adequate description and solution methods to each component. Explicit cosimulation, in particular, is frequently used when efficient code execution is a requirement, for instance in realtime setups. Using explicit schemes, however, can lead to the introduction of energy artifacts at the discretetime interface between subsystems. The resulting energy errors deteriorate the accuracy of the cosimulation results and may in some cases develop into the instability of the numerical integration process. This paper discusses the factors that influence the severity of the energy errors generated at the interface in explicit cosimulation applications, and presents a monitoring and correction methodology to detect and remove them. The method uses only the information carried by the variables exchanged between the subsystems and the cosimulation manager. The performance of this energycorrection technique was evaluated in multirate cosimulation of mechanical and multiphysics benchmark examples.
1 Introduction
Simulation has proven to be a valuable tool for the design, development and operation of engineering applications. Recent improvements in computing power and software tools have expanded its capabilities considerably, but expectations for its efficiency, the precision of the results that it delivers, and the complexity of the systems that it can successfully handle continue to grow as well. In the case of complex engineering applications such as automobiles, systemlevel simulation now needs to consider not only mechanical phenomena like suspension dynamics, but also processes of a different nature, such as hydraulics [5] or electronics [29], as well as the interactions between them. In many cases, it is possible to carry out the numerical integration of the dynamics of all these subsystems with a single, allencompassing solver, e.g., [32], following a monolithic approach. Cosimulation [12], or solver coupling, is an alternative solution that features a number of advantages.
In a cosimulation setup, the overall system is divided into subsystems and a dedicated simulation tool, or solver, is assigned to each of them. Each solver takes care of the numerical integration of its subsystem, which enables the use of tailored modelling and solution methods. It is also possible to use different computing platforms, both in terms of hardware and software, to run each integration process [18]. The exchange of information between subsystems is limited to discrete instants in time known as communication points; the interval between two consecutive communication points is a macrostep. Often, only a reduced set of coupling variables, i.e., inputs and outputs, is transferred at these instants. Between the communication points, the integration of the dynamics of each subsystem proceeds without additional data from its environment. The information about the subsystem internals is thus not disclosed to thirdparty software, which is an attractive feature from the point of view of preserving intellectual property rights in collaborative projects. In this regard, the seamless coupling of software tools from different vendors has been simplified by the introduction of interface standards like FMI [1]; in fact, in many industrial applications cosimulation subsystems can effectively be considered to behave as black boxes. Moreover, certain cosimulation schemes allow the distribution of the computational workload between several processing units, which can be used to parallelize code execution and improve efficiency [20].
Cosimulation setups can be built according to a wide variety of schemes and configuration options; the computational performance and accuracy of the results are considerably affected by this selection [24, 27]. Two major groups of coupling schedules can be identified: implicit and explicit ones. In implicit cosimulation schemes, macrosteps are repeated to improve the convergence of results. Explicit methods, conversely, do not perform rollback and macrosteps cannot be retaken once their integration has finalized. Implicit cosimulation methods are typically more stable and accurate than their explicit counterparts [17, 33]; nonetheless, noniterative algorithms are the option of choice in some cases. Efficient execution may be incompatible with iteration over macrosteps, for instance in applications in which the available time for computations is limited. In other cases, one or more subsystems do not allow rollback. Cyberphysical systems, in which physical components interact with virtual environments, are an example of an application subjected to these two constraints. On the one hand, all the subsystems in the setup must deliver realtime performance; depending on the complexity of the computational models used, some virtual components may not be able to meet this requirement if they have to iterate over their macrosteps. Besides, it is generally not possible to revert the state of a physical component to a previous instant in time.
Discretetime communication at the cosimulation interface has been identified as a source of discontinuities [10] and delays [7] in the inputs that the subsystems receive at the start of their macrosteps. These issues cause the cosimulation results to deviate from theoretically correct solutions [19], which translates into violations of the energy balance of the system [14], particularly when explicit schemes are used. This can make the simulation results unreliable and, in extreme cases, render the numerical integration unstable. Techniques like input extrapolation [23] are commonly used to alleviate this problem. Subsystem inputs can also be evaluated employing physicsbased predictions delivered by reduced subsystem models [25, 26], or approximated using directional derivatives that relate coupling variables and subsystem states [15, 16, 35]. Input prediction can deliver significant improvements with regard to cosimulation quality. It is, however, difficult to remove coupling errors completely, or even quantify them using this technique alone with explicit schemes.
Input correction is an alternative means of preventing the degradation of cosimulation results, based on modifying the subsystem inputs so that they comply with certain physicsbased criteria. This is the strategy followed in [8], which introduces a compensation signal to remove energy inconsistencies derived from input extrapolation. Another correction method was presented in [13], that enforces the conservation of the energy balance of the whole system by means of passivitybased control, assuming that information about the internal energy state of each subsystem is available. Control criteria can also be used to define similar correction algorithms [11]. An advantage of this strategy is that corrections are performed based on information about the system that can be used to evaluate cosimulation quality. Energybased indicators can be used to gather this information: The concepts of power residual and energy residual were introduced in [30]. They were initially used to define algorithms to control the macrostep size, but can also serve as a basis to develop methods to correct the system energy [31].
This paper puts forward an energymonitoring and inputcorrection method for explicit cosimulation that uses the information conveyed by the coupling variables exchanged between the subsystems. This algorithm can be used with Jacobi schemes in which the product of input and output variables has power units, e.g., in forcedisplacement coupling applications. The ability of the proposed approach to keep the cosimulation stable and accurate was evaluated by means of a set of benchmark problems, which included mechanical and multiphysics systems, both linear and nonlinear.
Sect. 2 in this paper provides an overview of the features and issues of explicit Jacobi cosimulation schemes. A discussion of the problems with energy conservation derived from the use of discrete cosimulation interfaces in these cases, together with the indicators used to monitor them, can be found in Sect. 3. A correction method for removing energy artifacts is presented here as well. Section 4 shows the application of the indicators and the energy correction algorithm to several explicit cosimulation scenarios. The results are discussed in Sect. 5; Sect. 6 summarizes the conclusions of the research.
2 Explicit Jacobi cosimulation schemes
Jacobi schemes are commonly used in explicit cosimulation. This arrangement allows for simultaneous integration of subsystems, a convenient feature for the parallelization of code execution, which becomes more complicated if sequential GaussSeidel schemes are adopted. Figure 1 illustrates the way in which the numerical integration proceeds between two communication points according to this scheme, for an application with two subsystems. At time \(t= t _{ k }\), the subsystems submit their outputs \(\mathbf { y } _{ k } ^{1}\) and \(\mathbf { y } _{ k } ^{2}\) to the cosimulation manager (step 1), which in turn evaluates and sends back the inputs \(\mathbf { u } _{ k } ^{1}\) and \(\mathbf { u } _{ k } ^{2}\) to them (step 2). At this point, both subsystems can independently carry out their integration from time \(t _{ k }\) to \(t _{ k + 1}\) (step 3). The process is restarted at time \(t _{ k + 1}\) when the new outputs \(\mathbf { y } _{ k + 1} ^{1}\) and \(\mathbf { y } _{ k + 1} ^{2}\) are forwarded to the manager (step 4).
Jacobi schemes can be singlerate if the communication stepsize is the same for all subsystems, or multirate if different stepsizes are used instead. The energy correction presented in this paper is evaluated differently in each case; direct feedthrough is another factor that needs to be taken into consideration as well.
The concept of direct feedthrough denotes the explicit dependence of the subsystem output from its input. Its presence in a cosimulation scheme may lead to the existence of algebraic loops, which often motivates the necessity to perform the initialization stage in an iterative way [2]. Besides, during runtime, direct feedthrough is associated with a deterioration of the convergence and stability properties of numerical integration [4].
Direct feedthrough motivates the need to extrapolate subsystem inputs in explicit singlerate Jacobi schemes, even when single step integrator formulas are used inside the subsystems. This extrapolation often becomes a source of coupling errors. The issue can be illustrated assuming that the dynamics of a given subsystem can be expressed using the discrete statespace notation as follows
where \(\mathbf { x } \) is the subsystem state, and \(\mathbf {A} \), \(\mathbf {B} \), \(\mathbf {C} \), and \(\mathbf {D} \) are the state, input, output and feedthrough matrices, respectively. In a singlerate Jacobi scheme like the one in Fig. 1, if explicit integration formulas are used in the subsystems, the evaluation of the state derivatives \(\dot { \mathbf { x } } \) in the macrostep \([ t _{ k }, t _{ k + 1}]\) only needs to be performed at instant \(t _{ k }\) to evaluate \(\mathbf { x } _{ k + 1}\). At this time, both the state \(\mathbf { x } _{ k }\) and the input \(\mathbf { u } _{ k }\) are known, and so the discretetime interface does not introduce any error in the computation of the system derivatives.
The subsystem output \(\mathbf { y } \), on the other hand, must be evaluated and returned at time \(t _{ k + 1}\) as
and this in principle requires input \(\mathbf { u } _{ k + 1}\), which is not available yet. If the subsystem has no direct feedthrough, i.e., \(\mathbf {D} = \mathbf {0} \), the output \(\mathbf { y } _{ k + 1}\) can be directly evaluated as
The outputs delivered by Eq. (3) only include the numerical error caused by the integration formula used to update the state between \(t _{ k }\) and \(t _{ k + 1}\). On the contrary, a subsystem with direct feedthrough will feature \(\mathbf {D} \neq \mathbf {0} \) and an approximation \(\tilde {\mathbf { u }} _{ k + 1}\) of the inputs at \(t _{ k + 1}\) is necessary during output evaluation,
The approximated input \(\tilde {\mathbf { u }} _{ k + 1}\) can be obtained in different ways, e.g., via polynomial extrapolation, but will generally not equal \(\mathbf { u } _{ k + 1}\). This approximation introduces an additional error in the evaluation of the system dynamics, which deteriorates the accuracy of the cosimulation results.
However, there exist several scenarios in which input extrapolation introduces additional sources of error into the cosimulation process, even if the subsystems are not subjected to direct feedthrough:

Implicit integration formulas require the derivative \(\dot { \mathbf { x } } _{ k + 1}\) in the evaluation of \(\mathbf { x } _{ k + 1}\). Input approximation is necessary for this, as evidenced by Eq. (1).

Multirate cosimulation schemes make it necessary to evaluate the derivatives of the subsystem state at intermediate points between the communication points \(t _{ k }\) and \(t _{ k + 1}\), giving rise to the need for input extrapolation as well, except in the rather infrequent case in which \(\mathbf {B} = \mathbf {0} \).

In the case of nonlinear systems, matrices \(\mathbf {A} \), \(\mathbf {B} \), \(\mathbf {C} \), and \(\mathbf {D} \) may be functions of the state \(\mathbf { x } \) and the input \(\mathbf { u } \).
The abovementioned circumstances have to be taken into account for a correct evaluation of the energy imbalance at the coupling interface, besides other errors, e.g., those that stem from the integration process, which also introduce inaccuracies in the results, although their impact is frequently less critical than that of input extrapolation [11]. These issues will be addressed in Sect. 3, which discusses how to calculate the energy errors introduced by cosimulation and puts forward a method to eliminate this artificial modification of the system energy.
3 Energybased indicators for explicit cosimulation
Determining the stability of generic cosimulation applications is a complicated task. Quantifying how much they deviate from theoretically correct results is even more challenging. Because such reference solutions are often not available for comparison, indicators have been proposed and employed to convey this information.
If the integration formulas used inside the subsystems are known, it is possible to evaluate the stability of the coupling scheme by its spectral radius [10, 33, 36]. A measure of stability, however, does not include all relevant information about the accuracy of the results. The convergence of particular coupling methods and schemes can be evaluated, e.g., [4, 35], but it is not straightforward to use these to quantify the actual error introduced by a particular simulation case, although error bounds can be defined in particular cases [3]. Some methods to quantify the errors introduced in the cosimulation results have been put forward nonetheless. In general, these require some information about the subsystem internals, either directly obtained [20] or provided by the directional derivatives of the subsystems [15].
In many explicit cosimulation applications, the subsystems behave as blackboxes and so the information about their internals is not accessible to the manager. In these cases, it is useful to introduce indicators that can be evaluated using only the information carried by the coupling variables. This is the case of the residual power and energy introduced in [30], defined for cosimulated systems in which the product of input and output variables has units of power. A method to adjust the macro stepsize of the setup was developed based on these, which was later [31] combined with an energy removal method similar to the one in [8] to improve the accuracy of the results. A similar energycorrection algorithm can be found in [13]; this one, however, requires information about the energy balance of the overall cosimulated system.
The power and energy residuals are used next as indicators for cosimulation quality and an interpretation of their physical meaning is provided. Then, an energycorrection method is developed and its performance is shown with a set of cosimulation benchmark problems.
3.1 Energybased indicators for cosimulation
The schematic of a monolithic solver for a system with three components is depicted in Fig. 2.
If the numerical integration errors are neglected, the overall energy balance for a mechanical system at any instant in time \(t\) is given by
where \(E \) is the mechanical energy, \(E _{0}\) is the initial energy level, and \(W _{\mathrm {nc}}\) is the work of nonconservative forces. Ideally, Eq. (5) must hold at all times. Thus, the power exchange between the three components in Fig. 2 must satisfy the conservation of energy, which is expressed as
where \(P _{Coi}\), \(i=1,2,\ldots , {n_{\textrm {co}}} \), is the power exchanged between each component and the rest of the system, and \({n_{\textrm {co}}} \) is the number of components in the system. Equation (6) means that, in the absence of external sources or sinks of energy, all the power flows without losses at the interface between the components. This monolithic scheme contains a single integrator, which takes care of the complete system and is allowed to access the internal information, e.g., the state, of each component at any point in time.
If a cosimulation scheme is used instead, as illustrated in Fig. 3, the power exchanged at the interface should also verify Eq. (6). Again, even if numerical integration errors are neglected, the system energy is modified by the cosimulation scheme due to the discretetime communication between the manager and the subsystems [13, 30]. The violation of the power conservation can be expressed by means of the residual power \(\delta P \) as
where \(P _{SSi}\), \(i=1,2,\ldots , {n_{\textrm {ss}}} \), is the power exchanged between a subsystem and the cosimulation manager through the interface, and \({n_{\textrm {ss}}} \) is the number of subsystems.
The residual power \(\delta P \) can be used to assess cosimulation quality when the product of subsystem inputs and outputs has units of power. Its evaluation does not require access to the subsystem internals, but only the information carried by the coupling variables. This is the case for a wide range of practical cosimulation setups, for instance, forcedisplacement or torqueangular speed couplings in mechanical systems or voltagecurrent exchanges in electric or electronic circuits.
For some selections of coupling variables, the inputoutput product could also deliver other units related to power exchange, e.g., \({\; \mathrm {W/m^{2}}}\) s. In these situations, it would also be possible to evaluate the residual power, with the additional requirement that the cosimulation manager would need to know some physical properties of the subsystems.
Assuming that the product of the coupling variables in a certain cosimulation configuration has power units, it is possible to evaluate the residual power \(\delta P \) for singlerate Jacobi schemes with two subsystems as [30]
where \(\mathbf { u } \) and \(\mathbf { y } \) are the arrays of inputs and outputs handled by the cosimulation manager. Ideally, Eq. (8) should equal zero if no errors existed at the coupling interface.
Figure 4 shows power exchanges in a singlerate explicit Jacobi schedule. The superscripts \(() ^{1}\) and \(() ^{2}\) denote the first and second subsystems, whereas subscripts \(() _{ k }\) and \(() _{ k + 1}\) refer to the timestep indexes \(t _{ k }\) and \(t _{ k + 1}\), respectively. This figure also illustrates the issue with the evaluation of subsystem outputs at time \(t _{ k + 1}\). As already discussed in Sect. 2, the input \(\mathbf { u } _{ k + 1}\) is necessary to determine outputs \(\mathbf { y } _{ k + 1}\) in subsystems with direct feedthrough. In fact, \(\mathbf { u } _{ k + 1}\) is also needed to complete the integration of the dynamics from \(t _{ k }\) to \(t _{ k + 1}\) if implicit integrators like the Newmark family [22] are employed. For simplicity, the discussion here is limited to subsystems with explicit integrators, e.g., the symplectic Euler formula.
Several ways to approximate \(\mathbf { u } _{ k + 1}\) can be used; polynomial extrapolation is widely used for this purpose [9, 23, 27], especially loworder methods such as constant or zeroorder hold (ZOH), linear or firstorder hold (FOH), and quadratic or secondorder hold (SOH), as shown in Fig. 5. Accordingly, outputs \(\mathbf { y } _{ k + 1}\) are evaluated with an approximation \(\tilde {\mathbf { u }} _{ k + 1}\) instead of the input value \(\mathbf { u } _{ k + 1}\) and the residual power becomes
which, in general, will no longer satisfy \(\delta P = 0\).
The evaluation of residual power corresponds to the cosimulation manager; in general, it cannot be assumed that all subsystems evaluate their residual power and include it among their coupling variables. This means that some assumption has to be made about the method that each subsystem uses to approximate its inputs \(\mathbf { u } _{ k + 1}\). Here we consider that \(\tilde {\mathbf { u }} _{ k + 1}= \mathbf { u } _{ k }\) in the evaluation of \(\delta P \) in Eq. (9).
The definition of the indicator in Eq. (9) can be generalized to Jacobi multirate coupling schemes as well [30].
The residual power \(\delta P \) can be used to keep track of the stability and accuracy of the integration process as the simulation proceeds. The indicator values can, for example, oscillate within an admissible range, which will be dependent on the application. In such cases, the cosimulated dynamics can be considered reasonably accurate. Conversely, the presence of sudden changes in the indicator can be a symptom of unstable behaviour and degraded accuracy of the results.
Moreover, the residual power \(\delta P \) is related to the accumulation of errors in the energy balance of the system. The deviations from the exact solution caused by input extrapolation will introduce an error in the balance in Eq. (5),
where \(\varsigma \left ( t\right ) \) is the energy deviation at time \(t\) caused by the discretetime interface.
The existence of a relation between the energy error \(\varsigma \) and the residual power \(\delta P \) was confirmed in [27]. The error \(\varsigma \), however, cannot be directly obtained by integrating the \(\delta P \) over time. As a matter of fact, depending on the selected coupling scheme and the properties of the subsystems, it is possible to approximate it with a fraction \(\mu \in[0,1]\) of this integral. For example, for the first cosimulation step, the violation of the energy balance is given by
As integration proceeds over time, however, the relation given by \(\mu \) may lose accuracy because the accumulation and propagation of errors brings the numerical integration further away from the theoretical solution. For this reason, the integral over time of the residual power, i.e., the residual energy \(\delta E \left ( t\right ) \), does not directly provide the value of the energy error \(\varsigma \left ( t\right ) \) in the system,
and the use of expression (11) must be limited to the calculation of the energy error introduced by the cosimulation environment between two consecutive communication points. A method for determining the value of \(\mu \) is discussed in Sect. 3.3.
3.2 Energy correction
The information conveyed by the residual power \(\delta P \) can be used to develop an energycorrection method aimed at removing the energy deviations derived from input extrapolation. The presented method is similar to the energy correction solution described in [13], in the sense that its objective is enforcing the satisfaction of the energy balance of the complete system, dissipating excess energy or introducing it back into the system if it has been lost, as an adaptive damping element would do. A major difference is that this method does not require the subsystems to provide information about their internal energy state, which is not possible in a number of practical applications. The cosimulation macro stepsize is not modified.
As mentioned in Sect. 3.1, the integral of the residual power \(\delta P \) over time does not correspond exactly to the accumulated energy error \(\varsigma \) of the overall system and needs to be corrected by means of a coefficient \(\mu \). The energy error between two consecutive communication points can then be approximated as
where \(\varsigma _{ k + 1}\) denotes the energy error added to the system between communication points \(t _{ k }\) and \(t _{ k + 1}\), and \(\delta E _{ k + 1}\) is the residual energy evaluated at time \(t _{ k + 1}\). Ideally, the energy error in Eq. (13) should be removed by the corrective action acting during the following macro stepsize, from \(t _{ k + 1}\) to \(t _{ k + 2}\), in order to readjust the energy level of the system. The nature of this action depends on the kinds of coupling variables being exchanged at the interface; in a forcedisplacement configuration, for instance, it is a force.
Figure 6 conceptually illustrates the correction scheme used. In this cosimulation graph, the output of subsystem 1, \(\mathbf {f} \), is an action, e.g., a force or torque. The output of subsystem 2, \(\mathbf {v} \), in turn, represents a kinematic quantity such as a velocity. It will be assumed in the following that the product \(\mathbf {f} ^{\mathrm {T}} \mathbf {v} \) has power units. In general, subsystems that return actions are more likely to present direct feedthrough, as they often need to know the position or velocity of other subsystems to evaluate their outputs. The proposed correction method modifies the action \(\mathbf {f} \) sent to the second subsystem with an additional term \({ \mathbf {f} } ^{\mathrm {corr}}\) that corrects the energy error from the previous macro step,
The action correction in Eq. (14) can be compared to the introduction of adaptive damping at the cosimulation interface. The applicability of Eq. (14) is limited by several facts. First, values of the coupling variable \(\mathbf {v} \) that approach zero will result in large values of the correction term \({ \mathbf {f} } ^{\mathrm {corr}}\), which can cause impulsive behaviour and destabilize the system dynamics. The maximum admissible correction must be limited in practical applications. As a consequence, the possibility of removing energy errors is limited at these points. It must be noted that the force limit used is problemdependent and depends on the nature and timescale of the subsystem dynamics. In this work it was set to a percentage of the uncorrected coupling force.
Additionally, the correction term \({ \mathbf {f} } ^{\mathrm {corr}}\) is calculated using the values of energy and velocity obtained at the end of the macro step \([ t _{ k }, t _{ k + 1}]\), but the correction action \({ \mathbf {f} } ^{\mathrm {corr}} _{ k + 1}\) has to be applied during the next macro step \([ t _{ k + 1}, t _{ k + 2}]\), at which the value \(\mathbf {v} _{ k + 1}\) used to evaluate the correction is unlikely to remain constant. For this reason, the energy correction resulting from the introduction of \({ \mathbf {f} } ^{\mathrm {corr}} _{ k + 1}\) will not match exactly the value of \(\varsigma _{ k + 1}\) that it intended to remove, but a difference \(U _{ k + 1}\) will exist between them. A possible way to approximate this value is to evaluate the difference between the intended correction and the one that would result if the coupling velocity at time \(t _{ k + 2}\) were used
The term \(U _{ k + 1}\) in Eq. (15) can be seen as the energy deviation that the method has been unable to correct during the last macro step. The accumulation of these terms over time leads to an uncorrected energy deviation \(\mathcal {U} \) in the overall system dynamics
An extra term can be introduced in the correction algorithm in Eq. (14), to remove the accumulated energy difference \(\mathcal {U} _{ k }\)
where \(\nu \) is a weight coefficient similar to an integral gain in the range \([0, 1]\), where 0 stands for no integral correction and 1 would enforce the removal, in theory, of all the accumulated energy error \(\mathcal {U} _{ k }\) in the next macro step.
It is also worth mentioning that in cosimulation environments that include physical components, the introduction of a correction term \({ \mathbf {f} } ^{\mathrm {corr}}\) has to be compatible with the nature of actuators used in the system. Additional improvements could be performed on the method, such as using other integration formulas in Eq. (13) or conducting some prediction of the velocity during the next communication step when calculating the correction action in Eq. (14).
3.3 Evaluation of correction coefficients
The indicators and the correction method presented in Sects. 3.1 and 3.2 require the determination of the fraction \(\mu \) of the residual power \(\delta P \) that indicates the spurious energy introduced into the system at the cosimulation interface.
The preliminary evaluation of the \(\mu \) coefficient was performed with the help of a twodegreeoffreedom linear oscillator, shown in Fig. 7, a system widely used in the cosimulation literature as benchmark problem, e.g., [12–14, 33].
The double linear oscillator is a mechanical system composed of two masses (\(m _{1}\) and \(m _{2}\)) connected to each other and to the ground by means of dampers (\(c _{1}\), \(c _{2}\) and \(c _{\mathrm{c}}\)) and springs (\(\kappa _{1}\), \(\kappa _{2}\), and \(\kappa _{\mathrm{c}}\)) with constant coefficients. The initial system displacements were set to \(\eta _{1}(0) = \eta _{2}(0) = 0{\; \mathrm {m}}\); the spring forces are zero in this configuration. The initial system velocities were \(\dot {\eta } _{1}(0) = 100{\; \mathrm {m/s}}\) and \(\dot {\eta } _{2}(0) = 100{\; \mathrm {m/s}}\), a set of values previously used in the literature, e.g., [13, 34, 35]. Several simulation cases, summarized in Table 1, were defined to verify the system behaviour for different sets of physical parameters.
Figure 8 shows the linear oscillator split into two subsystems in a forcedisplacement cosimulation configuration. The subsystem \(\mathcal{M}_{1}\) includes the coupling springdamper system and returns the force \(\mathbf { f } \) that it exerts on the second subsystem, \(\mathcal{M}_{2}\). This, in turn, takes care of the integration of the motion of \(m _{2}\) and delivers its displacement \(\eta _{2}\) and velocity \(\dot {\eta } _{2}\), denoted here as the array \(\mathbf { x } _{2}\).
Only the subsystem \(\mathcal{M}_{1}\) is subjected to direct feedthrough (\(\mathbf {D} _{\mathcal{M}_{1}} \neq \mathbf {0} \)), and requires to know its inputs to evaluate its outputs at any given instant in time. In this scheme, the inputs \(\mathbf { u } \) and outputs \(\mathbf { y } \) for each subsystem are given by
where the superscript \(() ^{*}\) points out that the input of a subsystem might not necessarily be equal to the output of the other one, because the cosimulation manager may perform modifications in their values, e.g., by extrapolation.
The double linear oscillator can be used to illustrate the relation between the residual energy, \(\delta E \), and the actual energy error \(\varsigma \) accumulated during the cosimulation. Figures 9 and 10 compare the mechanical energy \(E \) of the double linear oscillator cosimulated following an explicit, singlerate Jacobi scheme with the mechanical energy obtained with its analytical solution, \(E ^{\mathrm{ref}}\). The shown results correspond to cases 1 and 2 in Table 1.
These figures confirm that the residual power \(\delta P \) and its integral \(\delta E \) indicate the deviation of the cosimulated system dynamics from its theoretical energy balance. In conservative systems, this results into an increase in mechanical energy. In systems with dissipation, part of the energy error modifies the mechanical energy, and part affects the work exerted by nonconservative forces. Figure 10 highlights two interesting facts. First, the results show that the cosimulation is stable, in the sense that the system energy decreases with time; however, they also evidence that it is not accurate, because the system energy does not follow that of the reference solution. Second, the presence of dissipation within the subsystems decreases the severity of the energy errors introduced by the timediscrete coupling interface, bringing the cosimulated solution closer to its theoretically correct counterpart than in the conservative case.
The linear oscillator example was used to determine the value of the fraction \(\mu \) required to satisfy Eq. (11) and, thus, achieve an accurate correction of the system energy by means of Eq. (14). This was carried out through the simulation of case 1 in Table 1. Different integration and communication stepsizes were used; in multirate cosimulation scenarios, ZOH, FOH and SOH input extrapolation orders were tested. The symplectic Euler formula was used as the integration method in all the tests. The values of \(\mu \) obtained this way were used as a reference for the simulation of the nonlinear benchmark examples presented in Sect. 4 and the rest of cases described in Table 1. For the linear oscillator, this coefficient was found to depend on the existence of direct feedthrough in the subsystems. In singlerate schemes, the coefficient \(\mu \) equals 0.5 if one of the subsystems is subjected to direct feedthrough [30]; its value is zero if both subsystems are free from this issue, because input extrapolation at the interface does not take place and energy correction is not necessary. On the other hand, in multirate cosimulation schemes, the ratio between the macro stepsizes and the extrapolation order at the cosimulation manager also needs to be considered. In these schemes, however, the evaluation of the coefficient \(\mu \) becomes more complicated and it is necessary to consider the extrapolation order and the ratio \(R \) between subsystem stepsizes
where \(h_{\textrm {df}} \) and \(h_{\textrm {ndf}} \) are the macro stepsizes of the subsystems with and without directfeedthrough, respectively. The value of the coefficient \(\mu \) can then be calculated using the following expressions for a twosubsystem cosimulation:

The subsystem with direct feedthrough has the largest stepsize: the corrective coefficient is obtained as
$$\begin{aligned} \mu _{\mathrm{ZOH}} &= \mu _{\mathrm{SR}} \; (1.5  \frac{0.5}{ R }) \\ \mu _{\mathrm{FOH}} &= \mu _{\mathrm{SR}} \end{aligned}$$(20) 
The subsystem with direct feedthrough has the smallest stepsize: the corrective coefficient is obtained as
$$\begin{aligned} \mu _{\mathrm{ZOH}} &= \mu _{\mathrm{SR}} \\ \mu _{\mathrm{FOH}} &= \mu _{\mathrm{SR}} \; R \end{aligned}$$(21)
Subscripts ZOH (zeroorder hold) and FOH (firstorder hold) denote the input extrapolation order used in the subsystem with the smallest stepsize.
Figures 11 and 12 show the values of \(\mu \) calculated with Eqs. (20) and (21) for different values of the stepsize ratio \(R \) and the two mentioned input extrapolation methods. Although it is not represented in the figure, the \(\mu \) for higher order extrapolation, e.g., quadratic or SOH, is the same as the one used for FOH.
4 Benchmark problems
The general validity of the correction coefficients \(\mu \) and the applicability of the energy monitoring and correction methods put forward in Sect. 3 were verified with a set of cosimulation benchmark problems, which included nonlinear and multiphysics examples. In all cases, it was required that the product of their coupling variables delivered power units. The benchmarks were tested using single and multirate cosimulation schemes.
4.1 Multiple linear oscillator
The first system used as a benchmark is a modification of the double linear oscillator described in Sect. 3.3, in which the first mass \(m _{1}\) has been divided into \(n1\) masses, thus turning the system into the \(n\)degreeoffreedom linear oscillator shown in Fig. 13.
The initial system displacements were set to \(0{\; \mathrm {m}}\) for all masses. The velocities of the first \(n1\) masses were set to \(100{\; \mathrm {m/s}}\), whereas the velocity of the \(n\)th mass was set to \(100{\; \mathrm {m/s}}\). Table 2 summarizes the system properties for all tested scenarios.
Figure 14 shows the oscillator arranged according to a forcedisplacement coupling scheme in which the subsystems \(\mathcal{M}_{1}\) to \(\mathcal{M}_{n1}\) contain the information about their coupling springdamper systems and evaluate the force at the interface. The algebraic sum of these coupling forces, here denoted as \(\mathbf { f } \), is passed as input to the subsystem \(\mathcal{M}_{n}\), which handles the integration of the dynamics equations of mass \(m_{n}\). This subsystem returns as output its position \(\eta _{n}\) and velocity \(\dot {\eta } _{n}\), denoted here as \(\mathbf { x } _{n}\). Direct feedthrough is found here in the first \(n1\) subsystems, which need their input \(\mathbf { x } _{n}\) to evaluate their coupling forces.
4.2 Double nonlinear oscillator
The second benchmark example is another variation of the linear oscillator described in Sect. 3.3, in which the three springs and dampers can be replaced by nonlinear counterparts, thus transforming it into a double nonlinear oscillator. The nonlinear forces that are considered in this benchmark include

Nonlinear spring forces that follow a square root function of spring elongation as
$$ f =  \kappa \; {\mathrm{sgn}}( \eta ) \; \sqrt{ \left  \eta \right  } $$(22)where \({\mathrm{sgn}}()\) is the sign function, \(\eta \) is the elongation of the spring, and \(\kappa \) is a constant stiffness coefficient.

Nonlinear damping forces that follow a quadratic function of the damper elongation rate \(\dot {\eta } \) as
$$ f =  c \; {\mathrm{sgn}}( \dot {\eta } ) \; \left  \dot {\eta } \right  ^{2} $$(23)where \(c \) is a constant damping coefficient.
This nonlinear oscillator was tested in two different scenarios. In the first case, it is considered that all spring forces are nonlinear, as described in Eq. (22), and that damping is not present. The second case uses the nonlinear damping from Eq. (23) and linear spring forces. Table 3 summarizes the parameters used in the two scenarios proposed for this benchmark. The initial mass displacements and velocities were the same as those used for the linear oscillator: \(\eta _{1}(0) = \eta _{2}(0) = 0{\; \mathrm {m}}\), \(\dot {\eta } _{1}(0) = 100{\; \mathrm {m/s}}\) and \(\dot {\eta } _{2}(0) = 100{\; \mathrm {m/s}}\).
4.3 Coupled pendula
The third benchmark problem is a twodegreeoffreedom mechanical system comprised of two rigid pendula connected to each other at their free tips by a spring and a damper, as shown in Fig. 15.
The pendula consist of point masses \(m _{1}\) and \(m _{2}\), located at points \(\mathsf {1}\) and \(\mathsf {2}\), connected to two massless rods of lengths \(l_{1}\) and \(l_{2}\) and pinned to the ground at points \(\mathsf {A}\) and \(\mathsf {B}\), respectively. Revolute joints exist at points \(\mathsf {A}\), \(\mathsf {B}\), \(\mathsf {1}\), and \(\mathsf {2}\), so that the rods and the springdamper system can rotate freely around them. Gravity acts along the negative \(y\) axis, \(g = 9.81{\; \mathrm {m/s^{2}}}\). The system parameters used in the simulations are detailed in Table 4. The initial angles between the rods and the negative \(y\) axis were set to \(\theta _{1}(0) = {\pi}/{4}{\; \mathrm {rad}}\) and \(\theta _{2}(0) = {\pi}/{18}{\; \mathrm {rad}}\); the initial velocities were set to zero. In this initial configuration, the springdamper is unloaded, thus its natural length \(l _{0}\) is approximately \(8.756{\; \mathrm {m}}\).
The system was divided following a forcedisplacement arrangement, shown in Fig. 16, for the cosimulation tests. The subsystem \(\mathcal{M}_{1}\) integrates the dynamics of the first pendulum and evaluates the coupling force \(\mathbf { f } \), which is passed on as input to the second subsystem. The subsystem \(\mathcal{M}_{2}\) in turn solves the motion of the second pendulum and delivers the global positions \((x_{2}, y_{2})\) and velocities \((\dot{x}_{2}, \dot{y}_{2})\) of its free tip \(\mathsf {2}\), grouped in the array \(\mathbf { x } _{2}\). The subsystem \(\mathcal{M}_{1}\) is subjected to direct feedthrough. The corresponding coupling variables are
where again the superscript \(() ^{*}\) denotes the modifications applied to the coupling variables by the cosimulation manager. The term \(\mathbf { f } \) contains the \(x\) and \(y\) components of the coupling force exerted by the springdamper.
In order to solve the equations of motion of the pendulum within each subsystem, different multibody system dynamics formulations could be used. Here, an augmented Lagrangian algorithm [6] was used in combination with the symplectic Euler integration formula. The dynamics of each subsystem is formulated as a system of differential algebraic equations (DAEs), which introduces a difference with respect to the oscillators presented so far, in which the dynamics was described by ordinary differential equations (ODEs).
4.4 Hydraulic crane
The last benchmark is a hydraulic crane that consists of a twodegreeoffreedom multiphysics system composed of a twolink robotic arm actuated by a hydraulic cylinder, as shown in Fig. 17. A similar system was initially described in [21], and modified versions were later included as benchmarks in [25] and [28]. The example here is the same as used in [25], although a different manoeuvre is performed in the numerical tests.
This example is a multiphysics assembly composed of two parts: a planar multibody system and a hydraulic circuit. The mechanical part consists of two rigid rods, connected to each other and to the ground by revolute joints, of lengths \(l _{1}\) and \(l _{2}\), respectively. Link 1 is pinned to the ground at point \(\mathsf {A}\) and has a uniformly distributed mass \(m _{1}\), whereas link 2 is massless. Two additional point masses \(m _{2}\) and \(m _{3}\) are located at \(\mathsf {Q}\) and \(\mathsf {P}\), respectively.
The system is actuated with a hydraulic piston, shown in Fig. 18, and moves under gravity effects (\(g=9.81 {\; \mathrm {m/s^{2}}}\) along the negative \(y\) axis). The internal pressures in the cylinder chambers are \(p _{1}\) and \(p _{2}\); the force exerted by the actuator is
where \(a_{p} \) is the cross section of the piston, \(c \) is the dissipation coefficient of the actuator, and \({s_{c}} \) is the actuator length, which varies during motion. The dynamics of the hydraulic system can be described by the following system of differential equations
where \(l _{3}\) is the cylinder length, \(l _{4}\) and \(l _{5}\) are the lengths of the chambers on each side of the piston, \(a_{i} \) and \(a_{o} \) are the areas of the valves that connect the cylinder chambers to the pump and the tank in the hydraulic system (not represented for simplicity), \(\rho \) represents the density of the fluid, and \(c_{d} \) stands for the discharge coefficient of the valves. The quantities \(p _{\text{P}}\) and \(p _{\text{T}}\) are the hydraulic pressures at the pump and tank, respectively. The coefficients \(\delta_{\text{P1}}\), \(\delta_{\text{P2}}\), \(\delta_{\text{T1}}\), and \(\delta_{\text{T2}}\) are zero when the quantity inside the square root that precedes them is negative, otherwise they are equal to one. The terms \(\beta _{1}\) and \(\beta _{2}\) stand for the bulk modulus in each cylinder chamber and are evaluated as a function of fluid pressure as
where \(a \) and \(b \) are scalar constants that are fluid properties. Assuming that the two cylinder chambers have equal volume at the starting time of the simulation, the chamber lengths \(l _{4}\) and \(l _{5}\) are given by
where \({s_{c}^{0}} \) is the initial length of the actuator. The valve areas \(a_{i} \) and \(a_{o} \) are obtained as a function of the externally controlled displacement \(\gamma \) of a spool, which is an input of the system
where \(\gamma \in \left [ 0,1 \right ] \). A more detailed description of the components of the hydraulic system, as well as the formulation used to model its behaviour can be found in [25]. The mechanical parameters used for this benchmark problem are summarized in Table 5.
At time \(t= 0\), the initial angles of links 1 and 2 are \(\theta _{1}(0) = {\pi}/{6}{\; \mathrm {rad}}\) and \(\theta _{2}(0) = {3\pi}/{2}{\; \mathrm {rad}}\). The system is initially in a state of static equilibrium. The force exerted by the actuator during the manoeuvre under study is regulated by the controlled valve, whose spool displacement \(\gamma \) follows a sinusoidal law with variable amplitude \(A \) and constant frequency \(\omega = 2{\; \mathrm {rad/s}}\),
The spool displacement can vary between a completely closed (\(\gamma = 0\)) and a completely open (\(\gamma =1\)) valve. Initially, the spool displacement was set to \(\gamma _{0} \approx0.45435\), which corresponds to the static equilibrium of the system.
The different time scales required for the accurate solution of the multibody and hydraulic dynamics make it convenient to use multirate schemes in the cosimulation of this system. The integration stepsize required for the mechanical subsystem is of the order of \(h_{\mathcal{M}} = 1{\; \mathrm {ms}}\), while the hydraulic commonly requires steps around \(h_{\mathcal{H}} = 0.1{\; \mathrm {ms}}\). The forcedisplacement coupling arrangement of the mechanical (ℳ) and hydraulic (ℋ) subsystems is shown in Fig. 19. The subsystem ℳ integrates the multibody system dynamics using as input the force \(\mathbf { f } \) exerted by the piston and returns as output \(\mathbf { x } _{\mathcal{M}}\) the displacement \({s_{c}} \) and rate \({\dot {s}_{c}} \) of the cylinder. The hydraulic subsystem ℋ, in turn, integrates its dynamics to obtain the pressures in the cylinder and evaluates the hydraulic force using Eq. (25). In this configuration, the hydraulic subsystem ℋ features direct feedthrough. Thus, the coupling variables for this example are
5 Results
The explicit cosimulation of the linear oscillator described in Sect. 3, as well as the rest of benchmark problems introduced in Sect. 4, were used to assess the ability of the residual power \(\delta P \) to indicate cosimulation quality. The numerical experiments were also used to verify the energy correction method discussed in Sect. 3.2. Both singlerate and multirate Jacobi schemes were evaluated, using the different sets of physical parameters referred in Tables 1–5. In all cases, the integration stepsizes of the subsystems matched the size of their communication step with the manager.
For each example, the residual power \(\delta P \) was monitored first during a cosimulation run without corrective actions. In a second round of simulations, the energy correction method was evaluated; in these, the coefficient \(\mu \) was adjusted using the relations obtained for the linear oscillator shown in Eqs. (20) and (21). In most examples it was necessary to introduce an additional correction by means of the removal of the accumulated energy difference, as done in Eq. (17), in order to deliver optimal results.
The results obtained with the double oscillator in Sect. 3.3 have confirmed that an almost exact removal of the excess energy is possible for this linear system. Figure 20 illustrates the effect of the energy correction in the multirate cosimulation of case 1 in Table 1. A correction coefficient \(\mu = 0.725\) was used, determined with Eq. (20) for a stepsize ratio \(R =10\) and ZOH extrapolation. The effect of the energy correction is reflected in the predicted system motion, shown in Fig. 21 for the same scenario. The increase in motion amplitude that resulted from the accumulation of spurious energy has been removed and the corrected solution matches with the analytical reference. It must be noted that a small delay with respect to the reference remained in the results; this effect has already been observed with other energybased correction algorithms [13].
The values of \(\mu \) calculated with Eqs. (20) and (21) were used as reference for the cosimulation of the rest of cases in Table 1 and the benchmark examples. Figure 22 shows the effect of the energy correction on the position \(\eta _{2}\) of the mass \(m_{2}\) of the double linear oscillator in case 7 with a corrective factor \(\mu = 0.5\), corresponding to singlerate cosimulation. Again, the motion amplitude matches that of the reference and a small delay can be observed in the results. Figure 23 compares the mechanical energy of the system with and without corrections. In this case, the mass \(m_{2}\) in the subsystem that receives the corrected force is an order of magnitude smaller than the one in the other subsystem \(m_{1}\). This affects negatively the ability of the method to bring the energy level down to that of the reference. More effective removal of the excess energy can be achieved by introducing the integral term in Eq. (17). Fig. 24 illustrates the need to set a limit on the maximum correction force exerted at the interface. Large values of correction can be detrimental to simulation accuracy, whereas insufficient modifications can fail to remove completely energy errors. In all problems used in this paper, the maximum admissible correction force has not exceeded 100% of the uncorrected coupling force. The combination of an appropriate integral factor \(\nu \) and a force limit is problemdependent, and the development of an approach to select both systematically remains an open task within this research.
The proposed correction approach can also be used when the cosimulation environment includes more than two systems. This is the case with the multiple linear oscillator in Sect. 4.1, in which the lefthand side consists of a set of \(n1\) subsystems that pass their coupling force to the manager. This example was simulated using \(n=3\) and \(n=5\) subsystems. Figures 25 and 26 confirm that it is possible to extend the use of the correction method to systems composed of more than two subsystems, taking into account the structure of the power bonds present between the subsystems. In this example, a single correction force was applied to subsystem \(n\), as shown in Fig. 14.
The cosimulation of the nonlinear systems in Sect. 4 required the use of an integral correction in most cases. Figure 27 shows the mechanical energy of the double nonlinear oscillator in Sect. 4.2 that results from the use of different correction factors \(\mu \) in the absence of the integral term \(\nu \). The energy of the uncorrected system (\(\mu =0\)) grows consistently with time. The factor that corresponds to singlerate cosimulation, \(\mu =0.5\), initially performs an acceptable control of the system energy, but eventually is not able to appropriately handle the nonlinear behaviour of the benchmark. Raising \(\mu \) above 0.5 results in overcorrections, rendering the system stable at the cost of dissipating too much energy.
The introduction of an integral term \(\nu \) resulted in the ability to remove energy deviations in spite of the nonlinear behaviour of the test problem, as confirmed by Fig. 28. Different values of the integral term were tested, with \(\nu =0.25\) being the optimal value to keep the cosimulation stable and accurate. The displacement \(\eta _{2}\) with this configuration is represented in Fig. 29. In case 2, in which nonlinear damping was present, the severity of the energy deviations with respect to the reference solution was reduced and it was possible to correct the motion without including integral terms, as shown in Fig. 30.
Numerical experiments with the nonlinear oscillator also confirmed the role of the residual power, \(\delta P \), as an indicator of cosimulation quality with nonlinear systems. Figure 31 shows the timehistory of the residual power during the uncorrected cosimulation of case 1. Although the values vary over time, the average of the signal is above zero, which indicates the consistent growth of the energy stored in the system that can be observed in Figs. 27 and 28.
Similar results regarding the ability of the proposed method to correct energy deviations in the system were obtained with the rest of cosimulation benchmarks. Figure 32 compares the \(x_{2}\) coordinate during the simulation of the coupled pendulum example with and without energy correction. Even though the deviation from the reference solution introduced by the cosimulation scheme was relatively small, the correction method was able to bring the results closer to the reference motion. Figure 33 illustrates the relevance of correction: the spurious energy introduced at the interface is small, but it accumulates steadily over time, degrading the quality of the results as the simulation proceeds. The residual power, shown in Fig. 34, indicates this energy behaviour: Its value, albeit small, remains above zero during motion, which corresponds to a slow but steady increase of the system energy. The proposed correction method removed this excess energy; it must be noted that an integral term was necessary to achieve adequate results.
The last benchmark example, the hydraulic crane, is a multiphysics system in which the mechanical and hydraulic subsystems present very different time scales. Such a problem benefits in terms of efficiency from the use of multirate cosimulation schemes. In this example, due to the significant dissipation present at the hydraulic cylinder, the deviations of the uncorrected cosimulated dynamics from the reference solution remained relatively small in terms of both motion and energy, as long as the macrosteps for each subsystems, \(h_{\mathcal{M}}\) and \(h_{\mathcal {H}}\), were kept small enough to ensure the stability of the integration. Figure 35 shows the results obtained with FOH extrapolation and \(h_{\mathcal{M}} = 10{\; \mathrm {ms}}\) and \(h_{\mathcal{H}} = 0.25{\; \mathrm {ms}}\), which correspond to a stepsize ratio \(R = 1/40\). In this configuration, the actuator displacement follows closely, within a 2\({\; \mathrm {mm}}\) margin, the reference motion. Unlike in the previous examples, the coupling errors at the discretetime interface have removed energy from the system, instead of increasing the energy level. The energy correction, nonetheless, improves again the accuracy of the results.
This example confirmed as well the usefulness of the residual power \(\delta P \) as an indicator of the cosimulation quality. Figure 36 shows the actuator length \({s_{c}} \) during the uncorrected cosimulation of the manoeuvre described by Eq. (33), using \(h_{\mathcal{M}} = 11.5{\; \mathrm {ms}}\), \(h_{\mathcal{H}} = 0.1{\; \mathrm {ms}}\) and FOH extrapolation. The cosimulated solution shows an oscillation between \(t=3{\; \mathrm {s}}\) and \(t=5{\; \mathrm {s}}\), approximately, which does not correspond to the actual system motion, as evidenced by the comparison with the reference solution. In a practical application, however, a reference solution will not be available in most cases and such a comparison would not be feasible, so that it would be unclear whether the oscillation is caused by a numerical artifact or if it actually corresponds to the true system behaviour. In fact, the cosimulation recovers from the oscillation and remains stable, partly because of the dissipative behaviour of the cylinder. However, the timehistory of the residual power contains a clear spike at the time during which the oscillation occurs, as can be seen in Fig. 37, which permits the identification of the oscillations as a result of numerical errors and thus labelling the simulation as unreliable.
It is important to mention that in all simulated examples, it was necessary to select an upper limit for the magnitude of the correction force that was applied to the subsystems via Eq. (14) or Eq. (17). Using these equations, the theoretical correction force \({ \mathbf {f} } ^{\mathrm {corr}}\) can grow indefinitely at points when the velocity used as the coupling variable, \(\mathbf {v} \), approaches zero. In all tested cases, the magnitude of the correction force \({ \mathbf {f} } ^{\mathrm {corr}}\) was kept below that of its uncorrected counterpart \(\mathbf {f} \) at each instant in time, although the stability and accuracy of the results could be improved in some scenarios by lowering this limit. The development of an automated way to tune beforehand these force caps, as well as the integral term gain \(\nu \), for each particular simulation scenario, remains a currently open line of research.
6 Conclusions
Explicit schemes are frequently used in cosimulation applications when efficient performance is a requirement. Among these, fixedstep Jacobi schedules enable the parallelization of solver execution and are particularly suitable for their integration in demanding environments, such as those that demand realtime behaviour. These explicit schemes, however, are known to suffer from stability and accuracy issues, especially when the subsystem dynamics are subjected to direct feedthrough. In these cases, it is necessary to assess cosimulation quality during runtime to ensure the reliability of the obtained results.
Residual power, an indicator that can be evaluated from the coupling variables in forcedisplacement cosimulation configurations, can be used to monitor the quality of numerical integration and enable actions to prevent the simulation from delivering incorrect results and becoming unstable in extreme cases. This magnitude was also used to develop an energycorrection method to remove the energy inconsistencies caused by the discretetime cosimulation interface. The method introduces a corrective action in the coupling variables and requires limited knowledge of the subsystem internals. Its behaviour was demonstrated in the numerical integration of a linear oscillator, a commonly used benchmark problem in the cosimulation literature, and further verified with a set of additional test examples. These included nonlinear and multiphysics systems; multirate cosimulation schemes were tested beside singlerate ones. The results showed that residual power can be used to gain insight into the cosimulation quality, and that the proposed energy correction method can be employed to remove the spurious variations in system energy caused by discontinuities and delays at discrete cosimulation interfaces.
References
FMI – Functional Mockup Interface (2021). https://fmistandard.org/
Andersson, C.: Methods and tools for cosimulation of dynamic systems with the Functional Mockup Interface. Ph.D. thesis, Lund University (2016)
Arnold, M.: How to verify worst case asymptotic error bounds for cosimulation algorithms. In: González, F., Cuadrado, J. (eds.) Proceedings of COSIM2021, an International Symposium on Cosimulation and Solver Coupling in Dynamics, Ferrol, Spain, pp. 48–49 (2021)
Arnold, M., Clauss, C., Schierz, T.: Error analysis and error estimates for cosimulation in FMI for model exchange and cosimulation v2.0. Arch. Mech. Eng. 60(1), 75–94 (2013). https://doi.org/10.2478/meceng20130005
Barbosa, T.P., da Silva, L.A.R., Pujatti, F.J.P., Gutiérrez, J.C.H.: Hydraulic hybrid passenger vehicle: Fuel savings possibilities. Mechanics Based Design of Structures and Machines, 1–19 (2020). https://doi.org/10.1080/15397734.2020.1714447, Early access
Bayo, E., García de Jalón, J., Serna, M.A.: A modified Lagrangian formulation for the dynamic analysis of constrained mechanical systems. Comput. Methods Appl. Mech. Eng. 71(2), 183–195 (1988). https://doi.org/10.1016/00457825(88)900850
Benedikt, M., Watzenig, D., Zehetner, J., Hofer, A.: Macrostepsize selection and monitoring of the coupling errof for weak coupled subsystems in the frequencydomain. In: International Conference on Computational Methods for Coupled Problems in Science and Engineering  Ibiza, Spain (2013)
Benedikt, M., Watzenig, D., Zehetner, J., Hofer, A.: A nearly energypreserving coupling element for holistic weakcoupled system cosimulations. In: NAFEMS World Congress 2013, Salzburg, Austria (2013)
Burger, M., Steidel, S.: Local extrapolation and linearimplicit stabilization in a parallel coupling scheme. In: IUTAM Symposium on SolverCoupling and CoSimulation, pp. 43–56. Springer, Berlin, Germany (2019). https://doi.org/10.1007/9783030148836_3
Busch, M.: Continuous approximation techniques for cosimulation methods: Analysis of numerical stability and local error. Z. Angew. Math. Mech. 96(9), 1061–1081 (2016). https://doi.org/10.1002/zamm.201500196
Chen, W., Ran, S., Wu, C., Jacobson, B.: Explicit parallel cosimulation approach: analysis and improved coupling method based on Hinfinity synthesis. Multibody Syst. Dyn. 52(3), 255–279 (2021). https://doi.org/10.1007/s1104402109785x
Gomes, C., Thule, C., Broman, D., Larsen, P.G., Vangheluwe, H.: Cosimulation: A survey. ACM Comput. Surv. 51(3), 1–33 (2018). https://doi.org/10.1145/3179993
González, F., Arbatani, S., Mohtat, A., Kövecses, J.: Energyleak monitoring and correction to enhance stability in the cosimulation of mechanical systems. Mech. Mach. Theory 131, 172–188 (2019). https://doi.org/10.1016/j.mechmachtheory.2018.09.007
González, F., Naya, M.Á., Luaces, A., González, M.: On the effect of multirate cosimulation techniques in the efficiency and accuracy of multibody system dynamics. Multibody Syst. Dyn. 25(4), 461–483 (2011). https://doi.org/10.1007/s1104401092347
Haid, T., Stettinger, G., Watzenig, D., Benedikt, M.: A modelbased corrector approach for explicit cosimulation using subspace identification. In: Proceedings of the 5th Joint International Conference on Multibody System Dynamics, Lisbon, Portugal (2018)
Kraft, J., Klimmek, S., Meyer, T., Schweizer, B.: Implicit cosimulation and solvercoupling: efficient calculation of interfaceJacobian and coupling sensitivities/gradients. J. Comput. Nonlinear Dyn. (2021). https://doi.org/10.1115/1.4051823 Accepted manuscript
Kübler, R., Schiehlen, W.: Modular simulation in multibody system dynamics. Multibody Syst. Dyn. 4(2/3), 107–127 (2000). https://doi.org/10.1023/a:1009810318420
Lacoursière, C., Härdin, T.: FMI Go! A simulation runtime environment with a client server architecture over multiple protocols. In: Proceedings of the 12th International Modelica Conference, Prague, Czech Republic (2017). https://doi.org/10.3384/ecp17132653
Li, P., Yuan, Q.: Influence of coupling approximation on the numerical stability of explicit cosimulation. J. Mech. Sci. Technol. 34(6), 2289–2298 (2020). https://doi.org/10.1007/s122060200504x
Meyer, T., Kraft, J., Schweizer, B.: Cosimulation: Error estimation and macrostep size control. J. Comput. Nonlinear Dyn. 16(4), 041002 (2021). https://doi.org/10.1115/1.4048944
Naya, M., Cuadrado, J., Dopico, D., Lugris, U.: An efficient unified method for the combined simulation of multibody and hydraulic dynamics: Comparison with simplified and cointegration approaches. Arch. Mech. Eng. 58(2), 223–243 (2011). https://doi.org/10.2478/v1018001100164
Newmark, N.M.: A method of computation for structural dynamics. J. Eng. Mech. Div., ASCE 85(EM3), 67–94 (1959). https://doi.org/10.1061/jmcea3.0000098
Oberschelp, O., Vöcking, H.: Multirate simulation of mechatronic systems. In: Proceedings of the IEEE International Conference on Mechatronics, 2004. IEEE, Istanbul, Turkey (2004). https://doi.org/10.1109/icmech.2004.1364473
Olivier, B., Verlinden, O., Kouroussis, G.: Effect of applied force cosimulation schemes on recoupled vehicle/track problems. Multibody Syst. Dyn. 50(4), 337–353 (2020). https://doi.org/10.1007/s11044020097488
Peiret, A., González, F., Kövecses, J., Teichmann, M.: Multibody system dynamics interface modelling for stable multirate cosimulation of multiphysics systems. Mech. Mach. Theory 127, 52–72 (2018). https://doi.org/10.1016/j.mechmachtheory.2018.04.016
Peiret, A., González, F., Kövecses, J., Teichmann, M.: Cosimulation of multibody systems with contact using reduced interface models. J. Comput. Nonlinear Dyn. 15(4), 041001 (2020). https://doi.org/10.1115/1.4046052
Rahikainen, J., González, F., Naya, M.Á.: An automated methodology to select functional cosimulation configurations. Multibody Syst. Dyn. 48(1), 79–103 (2020). https://doi.org/10.1007/s1104401909696y
Rahikainen, J., González, F., Naya, M.Á., Sopanen, J., Mikkola, A.: On the cosimulation of multibody systems and hydraulic dynamics. Multibody Syst. Dyn. 50(2), 143–167 (2020). https://doi.org/10.1007/s1104402009727z
Rodríguez, B., Sanjurjo, E., Tranchero, M., Romano, C., González, F.: Thermal parameter and state estimation for digital twins of epowertrain components. IEEE Access 9, 97384–97400 (2021). https://doi.org/10.1109/access.2021.3094312
Sadjina, S., Kyllingstad, L.T., Skjong, S., Pedersen, E.: Energy conservation and power bonds in cosimulations: noniterative adaptive step size control and error estimation. Eng. Comput. 33(3), 607–620 (2017). https://doi.org/10.1007/s0036601604928
Sadjina, S., Pedersen, E.: Energy conservation and coupling error reduction in noniterative cosimulations. Eng. Comput. 36(4), 1579–1587 (2020). https://doi.org/10.1007/s00366019007834
Samin, J.C., Brüls, O., Collard, J.F., Sass, L., Fisette, P.: Multiphysics modeling and optimization of mechatronic multibody systems. Multibody Syst. Dyn. 18(3), 345–373 (2007). https://doi.org/10.1007/s1104400790760
Schweizer, B., Li, P., Lu, D.: Explicit and implicit cosimulation methods: stability and convergence analysis for different solver coupling approaches. J. Comput. Nonlinear Dyn. 10(5), 051007 (2015). https://doi.org/10.1115/1.4028503
Schweizer, B., Lu, D.: Semiimplicit cosimulation approach for solver coupling. Arch. Appl. Mech. 84(12), 1739–1769 (2014). https://doi.org/10.1007/s0041901408835
Schweizer, B., Lu, D.: Stabilized index2 cosimulation approach for solver coupling with algebraic constraints. Multibody Syst. Dyn. 34(2), 129–161 (2015). https://doi.org/10.1007/s110440149422y
Zhang, R., Zhang, H., Zanoni, A., Wang, Q., Masarati, P.: A tight coupling scheme for smooth/nonsmooth multibody cosimulation of a particle damper. Mech. Mach. Theory 161, 104181 (2021). https://doi.org/10.1016/j.mechmachtheory.2020.104181
Acknowledgements
F. González acknowledges the support of the Ministry of Economy of Spain through the Ramón y Cajal research program, contract no. RYC201620222. This work was partially financed by the Government of Galicia under grants ED431B2016/031 and ED431F2021/04 and the Ministry of Economy of Spain through project TRA201786488R “Técnicas de cosimulación en tiempo real para bancos de ensayo en automoción”.
Funding
Open Access funding provided thanks to the CRUECSIC agreement with Springer Nature. Authors also acknowledge the University of A Coruña/CISUG for the funding of the open access charges.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
The authors declare that they have no conflict of interest.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Rodríguez, B., Rodríguez, A.J., Sputh, B. et al. Energybased monitoring and correction to enhance the accuracy and stability of explicit cosimulation. Multibody Syst Dyn 55, 103–136 (2022). https://doi.org/10.1007/s11044022098125
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11044022098125
Keywords
 Cosimulation
 Energy monitoring
 Energy correction
 Energybased indicators
 Adaptive damping
 Explicit Jacobi schemes