Where Science Meets Machine Learning
Bridging Science and Data
Modern AI systems are often built purely on data, yet many scientific and engineering problems are governed by well-established physical laws and domain expertise. Hybrid AI models aim to bridge this gap by combining domain knowledge—such as physics equations, conservation laws, or expert rules—with the flexibility of machine learning.
In AI-DAPT, hybrid models specifically address the limitations of both physics-based and purely data-driven approaches. Physics models are interpretable and reliable but can be computationally expensive or simplified, while data-driven models may lack physical consistency or require large datasets. Hybrid modelling combines the strengths of both, leading to more reliable predictions that respect physical laws, better performance in data-scarce or noisy settings, and increased interpretability compared to black-box models.
AI-DAPT Hybrid Model Approaches
AI-DAPT organizes hybrid modelling around two complementary strategies, depending on how science and machine learning interact.
In the science-complements-machine-learning approach, domain knowledge directly informs model design or training. Examples include Physics-Informed Neural Networks (PINNs), where governing equations are embedded in the loss function, or physics-guided feature engineering to improve input representations.
In the machine-learning-complements-science approach, a physics-based model or simulator provides the core predictions, while machine learning enhances them. Residual learning is currently one of the key approaches implemented in AI-DAPT:
- A simulator generates baseline predictions for a given system.
- A machine learning model is trained to predict the residuals—the differences between the simulator output and real-world observations.
- The combined result corrects errors from the initial model, improving accuracy without discarding scientific knowledge.
Residual learning leverages existing scientific simulators by learning a data-driven correction to their outputs, refining predictions in regimes where simplified modeling assumptions or missing physics lead to systematic errors. Scientific simulators can also be integrated into pipelines for other purposes, such as surrogate modeling to approximate expensive simulations or simulation-based inference methods for parameter estimation.
The Hybrid Model Engineering Engine (HMEE)
The Hybrid Model Engineering Engine (HMEE) turns hybrid AI modelling into a practical tool within AI-DAPT. Fully embedded in the AI pipeline, HMEE lets users define, train, and manage hybrid models via DAVE’s graphical interface. More than that, through residual learning it is possible for users to also
Key capabilities of HMEE include:
- A catalog of baseline and hybrid model templates that can be added to workflows via drag-and-drop.
- Flexible configuration of training parameters such as learning rates and loss functions.
- Interfaces to integrate custom simulators for residual learning.
- Automatic tracking and storage of trained models and results in the AI-DAPT catalog, enabling reuse and deployment.
By integrating residual learning, HMEE allows demonstrators to improve existing simulations with minimal extra computational cost or create and integrate their own, combining scientific grounding with the adaptability of machine learning. This makes hybrid AI modeling accessible, scalable, and ready for real-world application across healthcare, energy, robotics, and industrial domains.