GMO-Mat: A foundation model-based generative multi-objective optimization framework for materials discovery

Chemical Foundation Models (FM) have successfully supported the creation of state-of-the-art property predictors, which evidence the quality of the corresponding latent space representations created by these models. When coupled with decoder models, these representations enable the creation of FM-based generative algorithms, leveraging the quality of the corresponding foundation models. In this work, we describe GMO-MAT, an under-construction framework to support the creation of FM-based generative algorithms for materials discovery with support to multiple objectives and constraints derived from properties and structural design space specifications. Properties' objectives and constraints are assessed through predictors built on top of the corresponding foundation models, which could also inform latent space exploration strategies (e.g., prediction gradients). GMO-MAT aims to support the creation of generative workflows by combining multiple algorithms for diversification (i.e., sampling), intensification (i.e., local search), and orchestration, such as Markov Chain Monte Carlo, Reinforcement Learning, GFlowNets, Multi-Objective Gradient Descent, and Meta-heuristics. We present supporting preliminary results in the context of generating sustainable strong acids by optimizing four related properties (pKa, LogKow, Ready Biodegradability, and LD50) through a Multi-Objective Weighted Gradient Descent from examples of strong acids.