Entropy-constrained ML and residual data augmentation for chemical kinetics

Direct numerical simulation of turbulent reactive flows remains one of the most complex challenges in computational engineering. Solving the Navier-Stokes equations together with detailed chemical kinetics requires a computational cost that, in practice, limits their application to academic or validation problems. However, recent advances in artificial intelligence offer a promising avenue to drastically reduce that effort without sacrificing accuracy. In this context, the concept of surrogate models based on machine learning with thermodynamic constraints, combined with strategies for augmenting residual data, emerges. This approach not only speeds up calculations, but ensures that predictions respect fundamental physical principles such as the second principle of thermodynamics.

The central idea is to replace the direct calculation of detailed chemical sources with a neural network that predicts reaction rates from a reduced thermochemical state. In order for the model to be reliable during temporal integration, an explicit non-negative entropy constraint is incorporated, which directs the evolution of the system in physically permissible directions. This type of constraint, known as physics-constrained machine learning, prevents the surrogate from generating unstable or non-physical behaviors, one of the main problems of purely statistical models in fluid dynamics. In addition, the residual-based data augmentation strategy allows new input conditions to be explored without the need to run additional costly detailed chemistry simulations. Instead, synthetic samples are constructed from the original set, scaling the residual error in a controlled manner. This opens the door to extensive parametric studies at a fraction of the original cost.

To understand the practical impact, imagine a gas turbine combustion chamber or a compression-ignition engine. High-fidelity modeling of the interaction between turbulence and chemical reactions is essential to designing more efficient systems with lower emissions. Today, engineers rely on simplified combustion models, such as single-stage flames or reduced mechanisms, which often sacrifice precision. Surrogate models trained with thermodynamic constraints can operate at a speed between ten and a hundred times faster than that of the detailed calculation, maintaining errors below 5% in key quantities such as temperature, majority species or heat released. This allows them to be directly integrated into large-scale simulations, such as those carried out by automotive, aeronautics or energy companies.

The breakthrough is not limited to combustion. Any problem involving chemical kinetics coupled with fluid transport—for example, biomass pyrolysis, nanomaterial synthesis, or atmospheric chemistry—can benefit. The key is that the entropy constraint acts as a physical regularizer that keeps the model within the actual thermodynamic manifold, while the augmentation of residual data provides an efficient way to cover the space of operating conditions. This combination is especially relevant when you want to explore the effect of variations in inlet temperature, equivalence ratio, or pressure, without having to repeat complete reagent simulations.

From a business perspective, the adoption of these substitute models represents an opportunity to reduce development times and computing costs in R+D centers. At Q2BSTUDIO we understand that integrating artificial intelligence into simulation processes requires a careful approach: it's not just about training a network, but about ensuring that the resulting software is robust, scalable, and easy to maintain. That's why we offer bespoke applications that incorporate physically consistent machine learning models, tailored to each customer's specific needs. Whether a substitute for a detailed chemical mechanism, an AI-based agent-based simulation engine, or a real-time optimization system is needed, our team can design the solution.

Technological infrastructure also plays a critical role. Training these models typically requires large volumes of data and parallel computing power. The AWS and Azure cloud services we deploy provide elastic environments to run simulations and trainings without upfront investment in hardware. In addition, cybersecurity is critical when handling sensitive design data or intellectual property models; our cybersecurity solutions protect both data and AI pipelines. On the other hand, the monitoring and adjustment of these substitutes can benefit from interactive dashboards built with Power BI, integrating the results of the simulations into dashboards for decision-making.

An emerging trend is the use of autonomous AI agents that manage the entire cycle: from the generation of boundary conditions to the validation of the surrogate with thermodynamic constraints. These agents can run parametric sweeps, detect regions where the model loses accuracy, and request additional data through residual augmentation. At Q2BSTUDIO we develop enterprise AI that integrates with existing simulation platforms, enabling progressive adoption without completely replacing traditional workflows.

For R+D teams, having business intelligence services that translate the results of these simulations into performance indicators is an added value. For example, a surrogate combustion model can power a plant's digital twin, and the data is visualized by dashboards that alert on deviations in emissions or efficiency. All of this is supported by a bespoke software infrastructure that connects sensors, models and analysts.

Research on entropy constraints and augmentation of residual data is still in its early stages in the industrial field, but conceptual tests show very promising results. The ability to reduce computational cost by an order of magnitude, while maintaining physical fidelity, can transform the way combustion processes and chemical reactions are designed. The key to their mass adoption will be the availability of software tools that allow engineers not specialized in artificial intelligence to implement these models with ease.

At Q2BSTUDIO, we combine engineering, data science, and software development expertise to build solutions that accelerate simulation without compromising accuracy. Our approach to custom applications ensures that each model is trained to constraints appropriate to the physical problem, and deployed on cloud platforms ready to scale. If your company is looking to reduce simulation times in combustion, chemistry or reactor processes, we can help you design a replacement based on machine learning with physical guarantees. The future of high-fidelity simulation lies in integrating physics into machine learning, and that integration is precisely what we offer.