The integration of machine learning (ML) into the physical sciences is reshaping computational paradigms, offering the potential to accelerate demanding simulations such as computational fluid dynamics (CFD). Yet, persistent challenges in accuracy, generalization, and physical consistency hinder the practical deployment of ML models in scientific domains. To address these limitations and systematically benchmark progress, we organized the ML4CFD competition, centered on surrogate modeling for aerodynamic simulations over two-dimensional airfoils. The competition attracted over 240 teams, who were provided with a curated dataset generated via OpenFOAM and evaluated through a multi-criteria framework encompassing predictive accuracy, physical fidelity, computational efficiency, and out-of-distribution generalization. This retrospective analysis reviews the competition outcomes, highlighting several approaches that outperformed baselines under our global evaluation score. Notably, the top entry exceeded the performance of the original OpenFOAM solver on aggregate metrics, illustrating the promise of ML based surrogates to outperform traditional solvers under tailored criteria. However, this does not imply that the winning solution could replace the OpenFOAM solver or that it was overall superior, even for this specific task. Drawing from these results, we analyze the key design principles of top submissions, assess the robustness of our evaluation framework, and offer guidance for future scientific ML challenges.

The interaction between humans and AI in safety-critical systems presents a unique set of challenges that remain partially addressed by existing frameworks. These challenges stem from the complex interplay of requirements for transparency, trust, and explainability, coupled with the necessity for robust and safe decision-making. A framework that holistically integrates human and AI capabilities while addressing these concerns is notably required, bridging the critical gaps in designing, deploying, and maintaining safe and effective systems. This paper proposes a holistic conceptual framework for critical infrastructures by adopting an interdisciplinary approach. It integrates traditionally distinct fields such as mathematics, decision theory, computer science, philosophy, psychology, and cognitive engineering and draws on specialized engineering domains, particularly energy, mobility, and aeronautics. Its flexibility is further demonstrated through a case study on power grid management.

Physical simulations are at the core of many critical industrial systems. However, today's physical simulators have some limitations such as computation time, dealing with missing or uncertain data, or even non-convergence for some feasible cases. Recently, the use of data-driven approaches to learn complex physical simulations has been considered as a promising approach to address those issues. However, this comes often at the cost of some accuracy which may hinder the industrial use. To drive this new research topic towards a better real-world applicability, we propose a new benchmark suite "Learning Industrial Physical Simulations"(LIPS) to meet the need of developing efficient, industrial application-oriented, augmented simulators. To define how to assess such benchmark performance, we propose a set of four generic categories of criteria. The proposed benchmark suite is a modular and configurable framework that can deal with different physical problems. To demonstrate this ability, we propose in this paper to investigate two distinct use-cases with different physical simulations, namely: the power grid and the pneumatic. For each use case, several benchmarks are described and assessed with existing models. None of the models perform well under all expected criteria, inviting the community to develop new industry-applicable solutions and possibly showcase their performance publicly upon online LIPS instance on Codabench.

Parallel master-slave evolutionary algorithms easily lead to linear speedups in the case of a small number of nodes... and homogeneous computational costs of the evaluations. However, modern computer now routinely have several hundreds of nodes - and in many real-world applications in which fitness computation involves heavy numerical simulations, the computational costs of these simulations can greatly vary from one individual to the next. A simple answer to the latter problem is to use asynchronous steady-state reproduction schemes. But the resulting algorithms then differ from the original sequential version, with two consequences: First, the linear speedup does not hold any more; Second, the convergence might be hindered by the heterogeneity of the evaluation costs. The multi-objective optimization of a diesel engine is first presented, a real-world case study where evaluations are very heterogeneous in terms of CPU cost. Both the speedup of asynchronous parallel algorithms in case of large number of nodes, and their convergence toward the Pareto Front in case of heterogeneous computation times, are then experimentally analyzed on artificial test functions. An alternative selection scheme involving the computational cost of the fitness evaluation is then proposed, that counteracts the effects of heterogeneity on convergence toward the Pareto Front.

Master-slave parallelization of Evolutionary Algorithms (EAs) is straightforward, by distributing all fitness computations to slaves. The benefits of asynchronous steady state approaches are well-known when facing a possible heterogeneity among the evaluation costs in term of runtime, be they due to heterogeneous hardware or non-linear numerical simulations. However, when this heterogeneity depends on some characteristics of the individuals being evaluated, the search might be biased, and some regions of the search space poorly explored. Motivated by a real-world case study of multi-objective optimization problem the optimization of the combustion in a Diesel Engine the consequences of different components of heterogeneity in the evaluation costs on the convergence of two Evolutionary Multi-objective Optimization Algorithms are investigated on artificially-heterogeneous benchmark problems. In some cases, better spread of the population on the Pareto front seem to result from the interplay between the heterogeneity at hand and the evolutionary search.