Transparent open-box learning network provides auditable predictions
- 44 Downloads
The transparent open-box (TOB) learning network algorithm adds the useful dimensions to machine learning of auditability and interrogation of each prediction made. It achieves this by making available for instant inspection the exact calculations and relationships it applies to its prediction for each data record in a dataset. The algorithm is applied here to predict the pool boiling heat transfer coefficient (PBHTC) for alumina-water-based nanofluids from a large dataset (870 data records with PBHTC varying from 0.33 to 65.68 kW m−2 K−1). The dataset, compiled from published sources and listed in full in a supplementary file, involves four relatively easy-to-measure input variables (nanoparticle size, operating pressure, nanoparticle concentration in water, and excess temperature). These input variables involve highly fragmented and nonlinear relationships with each other and PBHTC. The TOB predicts PBHTC to high accuracy (TOB: RMSE = 1.27; R2 > 0.99) and its predictions compare favorable to the accuracy achieved by an ANN model applied to the same dataset (ANN: RMSE = 1.09; R2 > 0.9960). The TOB algorithm involves a two-stage routine (Stage 1 matching; Stage 2 optimization) with no overt or underlying correlations involved. It generates dependent-variable predictions with high degrees of accuracy for datasets with irregular, fragmented and nonlinear input variables representing complex chemical and physical systems. Its transparency also provides key insights into the underlying dataset. The TOB algorithm overcomes the black box tendencies of many machine-learning algorithms. This makes the TOB algorithm suitable for deployment in situations where each prediction needs to be verified and supported with the complete step-by-step underlying calculations involved.
KeywordsLearning network transparency Prediction performance Nanofluid pool boiling Avoiding correlations Auditing predictions
- 2.Heinert M. Artificial neural networks–how to open the black boxes. In: Proceedings of application of artificial intelligence in engineering geodesy (AIEG); 2008. p. 42–62.Google Scholar
- 8.Moreno G, Oldenburg SJ, You SM, Kim JH. Pool boiling heat transfer of alumina-water, zinc oxide-water and alumina-water + ethylene glycol nanofluids. In: ASME summer heat transfer conference collocated with the ASME Pacific Rim technical conference and exhibition on integration and packaging of MEMS, NEMS, and electronic systems; 2005. p. 625–32.Google Scholar
- 12.Kim JH, Kim KH, You SM. Pool boiling heat transfer in saturated nanofluids. In: ASME international mechanical engineering congress and exposition; 2004. p. 621–8.Google Scholar
- 19.Cieśliński J, Kaczmarczyk T. The effect of pressure on heat transfer during pool boiling of water-Al2O3 and water-Cu nanofluids on stainless steel smooth tube. Chem Proc Eng. 2011;32(4):321–32.Google Scholar
- 21.Frontline Solvers. Standard excel solver—limitations of nonlinear optimization. https://www.solver.com/standard-excel-solver-limitations-nonlinear-optimization. Accessed May 2018.
- 28.Bose NK, Liang P. Neural network fundamentals with graphs, algorithms, and applications. New York: McGraw Hill; 1996.Google Scholar