Machine learning models are flexible and powerful, but they often function as black boxes with high data requirements, and they do not provide insight into the underlying physical or data-generating phenomena. On the other hand, real-world data mainly come from sensors or manual measurements that can suffer from low quality or quantity, and an understanding of the underlying system is important for the interpretation of the measurements.
Hybrid AI encompasses a broad range of methods for combining domain knowledge with data-driven methods. It is sometimes also referred to as scientific machine learning or physics-informed machine learning.
Advantages of hybrid AI
A hybrid approach that combines physical models with machine learning have several advantages:
1) It preserves or embeds existing knowledge about the system, avoiding relearning what is already known.
2) It improves the flexibility of the model, making it easier to generalize models and adapt to changing conditions´
3) The knowledge may help compensate for inferior data quality, giving better all-round accuracy.
The SINTEF approaches
At SINTEF, we research and apply many different approaches to hybrid AI and its practical applications.
Soft-constrained deep learning models such as regularisation or physics-informed neural networks (PINNs) guide the learning through additional terms in the loss function. This is a flexible approach that works in many different applications, but it requires a pre-existing partial model that matches the data well. In many cases this can be a good first step to compare existing assumptions to measurement data.
Hard-constrained deep learning models implement physical or mathematical guarantees into the model architecture, e.g., energy-preserving pseudo-Hamiltonian neural networks for ordinary and partial differential equations.
Learn analytical models of physical systems through use of machine learning, also known as system identification, whether it be in sparse systems or learning a system description with energy preservation. Inferring missing physics when first‐principles knowledge is incomplete or not consistent with the observed data.
Developing generative PIML methods for synthetic data/scenario generation, anomaly detection, imputation, and representation learning.
ML-in-the-loop focuses on applying machine learning methods in a control loop. In industry, most processes and systems are controlled to achieve stable and high production or to reach some goal, either through automated control systems or human operators in the loop. For complex systems, developing optimal, robust, and efficient controllers using traditional control theory is highly challenging. On the other hand, data collected from an active control loop is less varied and does not fully represent the underlying dynamics. A hybrid approach that combines existing domain knowledge and control theory with information in data may result in better estimation and control.
Large language models and foundation models can also be considered as interfaces and data interpreters for physical models and simulations.
Hybrid analytics is an active field of research, and SINTEF is participating at the forefront of practical applications.
Application areas
SINTEF has in-house competence in many domains, allowing for close collaboration between domain experts and machine learning experts. We have tailored solutions for process industry control systems, energy market predictions, drilling risers, predictive maintenance of marine propulsion systems, among others.