COMPUTATIONAL RESEARCH in BOSTON and BEYOND (CRIBB)

Small molecule solubility is a critically important property which affects the efficiency, environmental impact, and phase behavior of synthetic processes. Experimental determination of solubility is a time- and resource-intensive process and existing methods for in silico estimation of solubility are limited by their generality, speed, and accuracy. This work presents two models derived from the fastprop and chemprop architectures and trained on BigSolDB which are capable of predicting solubility at arbitrary temperatures for any small molecule in organic solvent. Both extrapolate to unseen solutes 2-3 times more accurately than the current state-of-the-art model and we demonstrate that they are approaching the aleatoric limit (0.5-1 logS), suggesting that further improvements in prediction accuracy require more accurate datasets. These models, collectively referred to as fastsolv, are open source, freely accessible via a Python package and web interface, highly reproducible, and up to 50 times faster than the next best alternative.

We demonstrate how E(3)-Equivariant Neural Networks (ENNs) can detect symmetry breaking in physical data and reveal hidden symmetry-implied information. Symmetry breaking, whether spontaneous (as in crystalline phase transitions) or induced by external forces, plays a crucial role in our understanding of physical systems. We show that an external input parameter to a fully equivariant model can be used to break symmetry in a physically interpretable way. We apply our approach to altermagnets, a new class of magnetic materials, in order to learn magnetic multipoles.

Large Language Models (LLMs) have reshaped generative AI, but fully fine-tuning these massive architectures is quite expensive in computational and communication resources. Low-Rank Adaptation (LoRA) partially mitigates these challenges, yet conventional LoRA often struggles to match the performance of full fine-tuning. In this talk, I introduce LoRA-SB (LoRA Silver Bullet), a novel approach that injects a constrained update space into LoRA’s framework, enabling optimal scaling for high-rank gradient directions that mimic full fine-tuning in a low-rank space, and meets the performance of full fine-tuning. We theoretically prove that our initialization strategy provides an optimal low-rank approximation of the initial gradient and preserves critical update directions throughout training. Extensive experiments on mathematical reasoning, commonsense inference, and language understanding tasks show that LoRA-SB exceeds the performance of standard LoRA while requiring 27–90× fewer trainable parameters and comprehensively outperforms LoRA-XS. Our findings demonstrate that it is not only possible but also highly effective to simulate full fine-tuning in low-rank subspaces, offering significant efficiency gains at no loss in accuracy. Additionally, we introduce Fed-SB, a federated extension of LoRA-SB that employs direct averaging of the small matrix R to guarantee exact updates and drastically reduce communication costs—independent of the number of clients—by up to 230×. Fed-SB further enhances privacy-utility-communication efficiency trade-offs by lowering noise requirements and avoiding noise amplification. Overall, it establishes a new Pareto frontier for efficient, scalable federated fine-tuning in both private and non-private settings.

While thermoset materials offer excellent performance, such as in renewable energy and aerospace applications, they are energy-intensive to make. Only a limited set of monomer chemistries have been identified as amenable to more energy efficient production via frontal polymerization. To learn the governing design principles and optimize for more sustainable high-performance polymer materials, it is critical to link the non-equilibrium process of front polymerization with a given monomer choice. We developed computational pipelines to learn how molecular structure informs the ability of a monomer resin to undergo energy-efficient manufacturing, specifically through frontal ring opening metathesis polymerization (FROMP). In our virtual automated workflow, we combinatorically enumerate candidate monomers and subsequently calculate key reaction parameters from first principles using density functional theory (DFT). We then integrate atomistic chemical insights into a mechanism-based reaction-diffusion model that simulates front propagations. This approach allows us to detail design rules for monomer choice and predict the suitability of a wide range of monomers for FROMP, purely from computation. Overall, our data-driven framework enables rapid processing and screening of new candidates, providing a foundation for in silico evaluation and discovery of FROMP amenable materials.