In collaboration with the Japanese company JSR, IBM has established a working group focused on developing new models, datasets, and benchmarks to tackle challenges in material science.

IBM Research has unveiled open-source foundation models designed for customization in various applications, including the search for improved battery materials and alternatives to hazardous PFAS chemicals.

These models utilize a mixture of experts (MoE) architecture, which effectively combines the strengths of different molecular models to enhance performance.

IBM researchers have pre-trained these models using various representation styles, such as SMILES-TED and SELFIES-TED, based on millions of validated samples.

These foundation models, pre-trained on extensive molecular databases, can efficiently screen millions of molecules to identify desirable properties while filtering out harmful ones.

The U.S. Environmental Protection Agency monitors nearly 800 toxic substances, which companies could replace with safer alternatives if available.

Additionally, IBM is collaborating with researchers through the AI Alliance to expedite the development of safer materials, focusing on critical areas such as reusable plastics and renewable energy.

Recent advancements in artificial intelligence provide innovative tools for identifying safer materials that prioritize human health and environmental safety.

However, the application of AI in chemistry faces significant challenges due to the complex three-dimensional structures of molecules, necessitating effective representation methods.

IBM plans to showcase these foundation models at the Association for the Advancement of Artificial Intelligence conference in February 2025, where they will also introduce new fusion techniques and models.

Different molecular representation formats, including SMILES, SELFIES, molecular graphs, and spectrograms, each have unique strengths and limitations that affect their application.

At the 2024 NeurIPS conference, IBM demonstrated that their MoE architecture outperformed leading models by employing a multi-view approach that integrates various data modalities.

The strong interest from the research community is evident, as the models have been downloaded over 100,000 times, according to Seiji Takeda from IBM Research.

The MoE architecture is adaptable to specific tasks, providing valuable insights into which data representations are most effective for different types of problems.