Evolutionary Physics-Informed Neural Networks for Fluid Mechanics

Overview

Fluid mechanics underpins a wide range of natural and engineered systems, from cardiovascular flows to wind energy and environmental transport. Many relevant flows are turbulent, multi-scale, and computationally demanding to model accurately.

This project develops evolutionary computation methodologies to make Physics-Informed Neural Networks (PINNs) more effective and efficient, with a primary focus on forward and inverse problems in fluid mechanics.

Why this project

PINNs combine sparse/incomplete data with governing equations through a composite loss. While promising— particularly for inverse problems—PINNs often face training instability and sensitivity to design decisions (loss balancing, activations, architecture, hyperparameters).

Evolutionary Computation (EC) provides robust strategies for complex optimization and has been successfully used to automate deep learning design and training. This project leverages EC to deliver generic, reusable strategies specifically tailored to PINNs.

Project goal

Main objective: deliver generic algorithmic tools and an open-source library that improve the training and design of PINNs for fluid mechanics (forward and inverse problems).

Core research directions

Automated loss design (balancing data and physics residuals)
Automated activation functions and training schemes
Architecture and hyperparameter selection as optimization
Distributed training paradigms
Training under limited data

Specific objectives

O1: Benchmarks in fluid mechanics (forward + inverse)
O2: Formalize PINN training/design as optimization
O3: Design EC algorithms for PINNs
O4: Open-source library + reproducibility

Team

Jamal Toutouh

Principal Investigator (PI)

University of Málaga (Spain)

Website Scholar ORCID

Sergio Nesmachnow

Researcher / Collaborator

Universidad de la República (Uruguay)

Scholar ORCID

Francisco Chicano

Researcher / Collaborator

Universidad de Málaga (Spain)

Web Scholar ORCID

Christian Cintrano

Researcher / Collaborator

Universidad de Málaga (Spain)

Web Scholar ORCID

Zakaria A. Dahi

Researcher / Collaborator

Universidad de Málaga (Spain)

Web Scholar ORCID

Martín Draper

Researcher / Collaborator

Universidad de la República (Uruguay)

Web Scholar

Published papers

Here you have published papers with the main outcomes related to the project.

Journal articles

Bove, M., López, B., Toutouh, J., Nesmachnow, S., Draper, M. (2023). Estimation of wind turbine wakes with generative-adversarial networks. Journal of Physics: Conference Series, 2505, 012053.
Toutouh, J., Nalluru, S., Hemberg, E., & O’Reilly, U. M. (2023). Semi-supervised generative adversarial networks with spatial coevolution for enhanced image generation and classification. Applied Soft Computing, 148, 110890.

Conference papers (CORE indexed)

Toutouh, J., Nalluru, S., Hemberg, E., & O'Reilly, U. M. (2023). Semi-supervised learning with coevolutionary generative adversarial networks. Proceedings of the Genetic and Evolutionary Computation Conference (GECCO), pp. 568–576.
Hemberg, E., O'Reilly, U. M., & Toutouh, J. (2024). Cooperative Coevolutionary Spatial Topologies for Autoencoder Training. Proceedings of the Genetic and Evolutionary Computation Conference (GECCO), pp. 331–339.
Sedeño, F., Toutouh, J., & Chicano, F. (2025). Generate more than one child in your co-evolutionary semi-supervised learning GAN. Applications of Evolutionary Computation (EvoStar), pp. 55–70. Springer Nature Switzerland.
Jorgensen, S., Hemberg, E., Toutouh, J., & O'Reilly, U. M. (2025). Guiding Evolutionary AutoEncoder Training with Activation-Based Pruning Operators. Proceedings of the Genetic and Evolutionary Computation Conference (GECCO), pp. 386–396.

Conference papers

Correa, J., Mignaco, J., Rey, G., Machín, B., Nesmachnow, S., & Toutouh, J. (2023). Multiobjective evolutionary search of the latent space of Generative Adversarial Networks for human face generation. Proceedings of the Companion Conference on Genetic and Evolutionary Computation (GECCO Companion), pp. 1768–1776.
Mignaco, J., Rey, G., Correa, J., Nesmachnow, S., & Toutouh, J. (2024). Empirical comparison of evolutionary approaches for searching the latent space of Generative Adversarial Networks for the human face generation problem. Proceedings of the Genetic and Evolutionary Computation Conference Companion (GECCO Companion), pp. 1631–1639.

Paper results: WakeGAN (Wake Conference 2023)

This section summarizes key results on GAN-based surrogates for wind turbine wakes, which motivates the broader agenda around physics-driven learning.

Problem setup

A conditional DCGAN is trained to predict the mean streamwise velocity at hub height in a wake region from an upstream inflow profile. Training data comes from LES–ALM simulations; local samples are interpolated to 64×64.

Quantitative results

Four setups vary the mean-velocity time window (TW1: 1000 steps, TW4: 4000 steps) and the MSE weighting factor (f_mse ∈ {0.0, 0.5}). Best test performance: TW4 with f_mse = 0.5.

Setup	Time window	f_mse	Test MRMSE (m/s)	Test FID	Time (s)
1	1000	0.0	0.39	0.46	895
2	1000	0.5	0.33	0.01	873
3	4000	0.0	0.22	1.10	444
4	4000	0.5	0.16	0.01	398

Interpretation: Shorter time windows show stronger spatial variability; adding the MSE term improves MRMSE.

Qualitative results

For TW4, the model captures wake region and main velocity features. Larger errors appear near wake edges; errors increase for TW1. Insert your paper figures below.

Sequential wind-farm prediction

When applying the GAN sequentially from upstream to downstream turbines, small errors can amplify downstream, and full-farm reconstruction can deteriorate with distance.

Practical takeaway

LES–ALM provides high fidelity but is expensive; the GAN surrogate captures key wake features and reduces runtime dramatically, highlighting the need for improved generalization.

Contact

PI: Jamal Toutouh — University of Málaga (Spain)

For collaboration, questions, or student projects, please contact the PI.

Send email GitHub