🙏 This is the joint work with my amazing colleagues @majhas.bsky.social and @nikhilshenoy.bsky.social. We also thank our fantastic collaborators Dominique Beaini, Stephan Thaler, and Hannes Stark for their invaluable contributions! A special mention to @valenceai.bsky.social for compute power💪

🌟6/6 ET-Flow sets a new benchmark for 3D molecular conformer generation, suggesting that equivariance is here to stay for molecular machine learning. Stop by our booth at East Exhibition hall at #NeurIPS 2024 on 11th December to connect, exchange ideas, and discuss potential collaborations! 🚀

🧬5/6 Results (2/2): Chemical Properties
ET-Flow doesn’t just generate molecules—it generates chemically and physically feasible molecules. This makes it highly impactful for downstream applications like drug discovery and materials science.

4/6 Results (1/2): Precision & Speed
ET-Flow achieves state-of-the-art precision in molecular conformer generation even with significantly few inference steps. While raw inference speed trails slightly, recent CUDA kernel optimizations for equivariant architectures will further boost performance.

🤔 3/6 Why Equivariance? ET-Flow incorporates equivariant design principles, resulting in a model with just 8.3M parameters, ~30x smaller than best performing non-equivariant baseline MCF-L (242M). This demonstrates how embedding symmetry as an inductive bias leads to efficiency.

2/6 We leverage harmonic prior and a rotation alignment, which further simplify learning conditional probability path between base and target distributions. Additionally, our stochastic sampling dynamically corrects the probability flow, enhancing the precision and accuracy of generated conformers.

💡1/6 Why Flow Matching? Flow Matching directly learns the probability flow between distributions, significantly reducing sampling complexity. This makes it especially suited for 3D molecular conformer generation, where precision and computational efficiency are key.

⭐ 6/6 ET-Flow sets a new benchmark for 3D molecular conformer generation, suggesting that equivariance is here to stay for molecular machine learning. Stop by at East Exhibition hall at #NeurIPS 2024 on 11th December to connect, exchange ideas, and discuss potential collaboration! 🚀

🧬5/6 Results (2/2): Chemical Properties
ET-Flow doesn’t just generate molecules—it generates chemically and physically feasible molecules. This makes it highly impactful for downstream applications like drug discovery and materials science.

📊4/6 Results(1/2): Precision & Speed
ET-Flow achieves state-of-the-art precision in molecular conformer generation with significantly few inference steps. While raw inference speed trails slightly, recent CUDA kernel optimizations for equivariant architectures will further boost performance.

🤔3/6 Why Equivariance? ET-Flow incorporates equivariant design principles, resulting in a model with just 8.3M parameters, ~30x smaller than best performing non-equivariant baseline MCF-L (242M). This demonstrates how embedding symmetry as an inductive bias leads to efficiency.

2/6 We leverage harmonic prior and a rotation alignment, which simplify learning conditional probability path between base and target distributions. Additionally, our stochastic sampling dynamically corrects the probability flow, enhancing the precision and accuracy of generated conformers.

💡1/6 Why Flow Matching(FM)? FM directly learns the probability flow between distributions, significantly reducing sampling complexity. This makes it especially suited for 3D molecular conformer generation, where precision and computational efficiency are key.