Foundation Models for Materials - FM4M

FM4M adopts a modular architecture designed for flexible extensibility. As illustrated in the figure below, it comprises both uni-modal and fused models. Each uni-modal model is pre-trained independently for its respective modality (e.g., SMILES), and users can access individual functionalities directly from the corresponding model directory (e.g., smi-ted/). Some of these uni-modal models can be "late-fused" using fusion algorithms, creating a more powerful multi-modal feature representation for downstream predictions. To simplify usage, we provide fm4m-kit, a wrapper that enables users to easily access the capabilities of all models through straightforward methods. These models are also available on Hugging Face, where they can be accessed via an intuitive and user-friendly GUI.