# The Well: a Large-Scale Collection of Diverse Physics Simulations for Machine Learning **Authors:** Ruben Ohana, Michael McCabe, Lucas Meyer, Rudy Morel, Fruzsina J. Agocs, Miguel Beneitez, Marsha Berger, Blakesley Burkhart, Stuart B. Dalziel, Drummond B. Fielding, Daniel Fortunato, Jared A. Goldberg, Keiya Hirashima, Yan-Fei Jiang, Rich R. Kerswell, Suryanarayana Maddu, Jonah Miller, Payel Mukhopadhyay, Stefan S. Nixon, Jeff Shen **arXiv:** [2412.00568](https://arxiv.org/abs/2412.00568) **Date:** 2024-11-30 **Summary page:** https://www.summarizepaper.com/en/paper/2412.00568v1/ ## Key points - Simulation plays a crucial role in predicting, optimizing, and inferring parameters for physical systems - Numerical methods are widely used due to the impracticality of finding analytical solutions for complex partial differential equations - Surrogate models have emerged as effective alternatives to numerical methods by capturing essential system features - Deep learning enhances surrogate modeling by providing faster and more accurate results across various fields - The Well dataset collection offers high-quality simulation data across diverse domains to drive innovation in developing efficient and accurate surrogate models - The Well serves as a valuable resource for machine learning researchers seeking challenging benchmarks to develop advanced data-driven surrogates - The Well provides 15 TB datasets covering various physical systems, each including temporally coarsened snapshots from simulations across different initial conditions or parameters - A unified PyTorch interface is offered by the Well for training and evaluating models, facilitating easy access and utilization of the data within machine learning workflows ## Summary , , , , In the realm of computational science and engineering, simulation plays a crucial role in predicting, optimizing, and inferring parameters for physical systems. However, finding analytical solutions for the complex partial differential equations that govern many phenomena is often impractical, leading to the widespread use of numerical methods. While these methods offer accuracy, they can be computationally expensive. To address this challenge, surrogate models have emerged as simplified yet effective alternatives that capture essential features of a system. Deep learning has shown promise in enhancing surrogate modeling by providing faster and more accurate results across various fields. Despite these advancements, the adoption of deep learning-based surrogates faces obstacles due to the mismatch between the complexity of real-world problems and available datasets. The introduction of example baselines showcases the unique challenges posed by the dynamic nature of the Well's datasets. Overall, the Well represents a significant step towards advancing machine learning-based surrogate modeling by offering researchers access to high-quality simulation data across various domains. The availability of this extensive dataset collection is expected to drive innovation in developing more efficient and accurate surrogate models for complex physical systems. The Well serves as a valuable resource for machine learning researchers seeking challenging benchmarks to develop advanced data-driven surrogates. By providing complex tasks at a manageable scale for modern machine learning techniques, the Well aims to facilitate the development of next-generation surrogate models that balance efficiency with accuracy. To bridge this gap, the Well is introduced as a comprehensive collection of 15 TB datasets derived from collaborations with domain experts and software developers. These datasets cover a wide range of physical systems such as biological processes, fluid dynamics, acoustic scattering, and astrophysical phenomena like supernova explosions. Each dataset includes temporally coarsened snapshots from simulations across different initial conditions or parameters to explore stability. In addition to offering diverse datasets, the Well also provides a unified PyTorch interface for training and evaluating models. This interface enables researchers to easily access and utilize the data within their machine learning workflows.