ExpressIF® is a modern implementation of symbolic AI, which is the knowledge-based form of AI. Developed by CEA-List, ExpressIF® models expert knowledge and can also automatically extract it from data. The knowledge produced can then be used for decision assistance or problem solving. The advantage of symbolic AI is that less data is needed to generalize and produce interpretable models, i.e., models that can be understood and checked by humans.
ExpressIF®, with its ability to combine expert knowledge and knowledge extracted from small datasets, is particularly useful for scientific research, especially in fields like materials science.
Frugal learning (learning on limited experimental data) and interaction with the user (a scientist in this case) are two of the main strengths of ExpressIF®. The development of ExpressIF® Materials takes these capabilities even further.
ExpressIF® Materials works by seeking causal relationships between production parameters and the properties of the materials produced. The relationships identified are then characterized by induced knowledge that effectively predicts the properties in question. The software can also find gradual relationships that facilitate the search for optimal parameters, which can be used to extract relevant information from experimental data. This transparency allows scientists to benefit from additional insights that are not available in “black box” AIs. ExpressIF® usually predicts better than other AI models (xgBoost, random forest, polynomial fitting, etc.) while providing access to information about how it arrived at a conclusion.
The current version of ExpressIF® has a new active learning algorithm that even recommends the next experiment to carry out. The knowledge is updated with each iteration and, as predictions of the material’s properties improve, the tool more rapidly arrives at the underlying production parameters. With each iteration the algorithm determines whether to explore a new possibility or to run an experiment similar to a previous one. Each decision is backed up with the underlying reasoning.
The tool offers a range of possibilities in terms of adapting to different kinds of data and to different use cases. Regardless of how and where it is used, it has the capacity to enhance the knowledge that can be extracted and boost performance.
Materials for solar panels, additive manufacturing, and corrosion-resistant parts. The tool can also be used for other experimental sciences like biology.
ExpressIF® is developed under the CEA’s cross-functional Materials and Processes competency plan.
O. Rousselle, J.-P. Poli and N. Ben Abdallah. Towards an interpretable fuzzy approach to experimental design. International Conference on Information Processing and Management of Uncertainty in Knowledge-Based Systems, 2024.
AI is particularly exciting for materials science due to the multidisciplinary nature of the field and its infinite potential to drive technological innovation and scientific research.