Physics-Guided AI for Turbocompressor Performance

DeepDynaSim Engineering & Consulting | MATLAB research workflow | turbo-pgml v35

Predict turbocompressor performance with a surrogate that combines regression models, thermodynamic relationships and residual correction. This repository includes training, inference, evaluation data, a saved model and an interactive MATLAB reporting dashboard.

Quick start · Method · Results · Technical report · Presentation

What you can explore

• Five outputs: efficiency ETA, total pressure ratio TPR, entropy-related ENT, ALPHA2 and VTH2.
• Physics-guided entropy coupling, residual learning and efficiency blending.
• Internal cross-validation, holdout metrics and evaluation on a supplied external set.
• Parity/residual plots, speed-bin diagnostics, CSV exports and PDF reporting.

This is a research software workflow for performance prediction. Its scope is the implemented surrogate and evaluation tools; application to a new compressor or operating domain requires validation with suitable data.

Repository map

All MATLAB files currently live in the repository root.

File	Purpose
run_v35_train_and_external.m	Train and evaluate a configured experiment
train_turbocompressor_model_v35.m	Fit the v35 surrogate
predict_turbocompressor_model_v35.m	Predict the five outputs
sanity_check_ent_dsfactor_v35.m	Inspect entropy/physics consistency
turbo_results_dashboard_ui_v35.m	Visualize and export evaluation results
cfg_params_and_run_v35.m	Existing configuration example; review its data and tuning settings
input.mat, output.mat, ext_input.mat, ext_output.mat	Included input/output matrices
turbo_model_v35.mat, turbo_external_report_v35.mat	Saved model and report

Quick start

Use MATLAB with Statistics and Machine Learning Toolbox for the regression models. Open MATLAB in the cloned repository directory. The dashboard uses MATLAB desktop UI components.

git clone https://github.com/DeepDynaSim/AI-Powered-Software-for-Turbo-Compressor-Design.git
cd AI-Powered-Software-for-Turbo-Compressor-Design

Inspect a saved model

addpath(pwd);
S = load("input.mat");
X = double(load_matrix_any(S, "input"));
Yhat = predict_turbocompressor_model_v35("turbo_model_v35.mat", X);
% Columns: ETA, TPR, ENT, ALPHA2, VTH2

Train and evaluate

Check the column meanings and working gas before running. Input and output rows must correspond; output columns are ordered as above. This example uses the included external filenames and explicitly disables external-label tuning.

addpath(pwd);
cfg = struct();
cfg.inputMat = "input.mat";
cfg.outputMat = "output.mat";
cfg.extInputMat = "ext_input.mat";
cfg.extOutputMat = "ext_output.mat";
cfg.speedCol = 1;
cfg.in23Col = 23;
cfg.gasName = "air";  % or "hydrogen", as appropriate for your data
cfg.modelOut = "turbo_model_v35.mat";
cfg.reportOut = "turbo_external_report_v35.mat";
cfg.etaTuneOnExternal = false;
run_v35_train_and_external(cfg);

Training writes the model/report paths above; use different filenames to retain the bundled artifacts. The existing configuration script uses the internal filenames for external evaluation and enables external-label tuning. Reusing training data is a demonstration, not independent validation. Any dataset used for tuning must be separated from the final untouched test set. Separate filenames alone do not establish independence.

Open the dashboard

turbo_results_dashboard_ui_v35( ...
    "modelPath", "turbo_model_v35.mat", ...
    "extInput", "ext_input.mat", ...
    "extOutput", "ext_output.mat", ...
    "predictFcn", "predict_turbocompressor_model_v35", ...
    "speedIdx", 1);

Method overview

1. TPR: select candidate regressors/transforms using cross-validation.
2. ETA: fit engineered features in logit space and apply polynomial bias correction.
3. ENT: couple efficiency, pressure ratio and the configured input column through the thermodynamic dsfactor relationship; learn a residual correction.
4. Efficiency blending: combine the data-driven estimate with the physics-derived estimate using the v35 blending/gating model.
5. ALPHA2 and VTH2: fit separate candidate regressors.

Results and reporting

The dashboard reports RMSE, MAE and R², input-shift summaries, parity plots and residuals. Exports include metrics.csv, shift.csv, eta_bins.csv, rows_preview.csv, output PNGs and a multipage PDF. Inspect ETA, TPR and ENT plots already in the repository.

Label training, internal holdout, tuning and final external results separately when reporting an experiment. Report the dataset provenance, split, MATLAB/toolbox versions, gas choice and column definitions with the metrics.

Troubleshooting

• Missing function: use the repository root as MATLAB current folder and run addpath(pwd). There is no src/ directory in this checkout.
• Unexpected entropy results: check in23Col, gasName and the third output column.
• Regression/toolbox errors: check Statistics and Machine Learning Toolbox availability.
• Different MAT field names: inspect the matrices and use load_matrix_any or adjust the loading step.

Maintainers, collaboration and license

Maintained by Caglar Uyulan, DeepDynaSim Engineering & Consulting Ltd. Co. Contributor: Sercan Acarer, SLC Fluidics Ltd.

DeepDynaSim is an engineering R&D and consulting company in İzmir, Türkiye, working on physics-guided AI, simulation and control. For engineering collaboration: support@deepdynasim.com. For reproducible bugs or technical questions, open an issue with the MATLAB version, toolbox list, configuration and a minimal example.

Code is distributed under the MIT license. When using this work academically, reference the repository commit or version and the applicable literature.

Related work: Satellite thermal preview control · Deep Lyapunov functions · Flexible-link manipulators

References (selected)

DOI, title and author metadata were checked against Crossref/DataCite during the documentation review. The compressor-cascade preprint and journal entry below are versions of the same work, not independent studies.

Karniadakis, G. E., Kevrekidis, I. G., Lu, L., Perdikaris, P., Wang, S., & Yang, L. (2021). Physics-informed machine learning. Nature Reviews Physics, 3, 422–440. https://doi.org/10.1038/s42254-021-00314-5

Karpatne, A., Atluri, G., Faghmous, J. H., Steinbach, M., Banerjee, A., Ganguly, A., … Kumar, V. (2017). Theory-guided data science: A new paradigm for scientific discovery from data. IEEE TKDE, 29(10), 2318–2331. https://doi.org/10.1109/TKDE.2017.2720168

Li, Z., Montomoli, F., & Sharma, S. (2024). Investigation of Compressor Cascade Flow Using Physics-Informed Neural Networks with Adaptive Learning Strategy. AIAA Journal. https://doi.org/10.2514/1.J063562

Li, Z., Montomoli, F., & Sharma, S. (2023). Investigation of Compressor Cascade Flow Using Physics-Informed Neural Networks with Adaptive Learning Strategy. Earlier arXiv version of the AIAA article above. https://doi.org/10.48550/arXiv.2308.04501

McClenny, L. D., & Braga-Neto, U. M. (2023). Self-adaptive physics-informed neural networks. Journal of Computational Physics, 474, 111722. https://doi.org/10.1016/j.jcp.2022.111722

Pawar, S., San, O., Aksoylu, B., Rasheed, A., & Kvamsdal, T. (2021). Physics guided machine learning using simplified theories. Physics of Fluids, 33(1), 011701. https://doi.org/10.1063/5.0038929

Pawar, S., San, O., Vedula, P., Rasheed, A., & Kvamsdal, T. (2022). Multi-fidelity information fusion with concatenated neural networks. Scientific Reports, 12, 5900. https://doi.org/10.1038/s41598-022-09938-8

Raissi, M., Perdikaris, P., & Karniadakis, G. E. (2019). Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear PDEs. Journal of Computational Physics, 378, 686–707. https://doi.org/10.1016/j.jcp.2018.10.045

Rathnakumar, R., Huang, J., Yan, H., & Liu, Y. (2024). Bayesian Entropy Neural Networks for Physics-Aware Prediction. arXiv. https://doi.org/10.48550/arXiv.2407.01015

Guo, M., Manzoni, A., Amendt, M., Conti, P., & Hesthaven, J. S. (2022). Multi-fidelity regression using artificial neural networks: Efficient approximation of parameter-dependent output quantities. Computer Methods in Applied Mechanics and Engineering, 389, 114378. https://doi.org/10.1016/j.cma.2021.114378

Willard, J., Jia, X., Xu, S., Steinbach, M., & Kumar, V. (2022). Integrating scientific knowledge with machine learning for engineering and environmental systems. ACM Computing Surveys. https://doi.org/10.1145/3514228