Multi-Source Data Unlocks Joint PDE Discovery with AI

Hao Xu, Siyu Lou, Yuntian Chen, Dongxiao Zhang· July 1, 2026 View original

Summary

This research introduces MCO-PDE, a competitive optimization framework that discovers shared partial differential equations (PDEs) from multiple datasets, even with limited observations. It uses independent neural surrogates for each source and a competitive weighting mechanism to aggregate a consensus global coefficient, successfully identifying governing laws from synthetic and real-world data.

A new framework called MCO-PDE (Multi-Competitive Optimization for PDEs) has been developed to automatically discover governing partial differential equations (PDEs) from diverse datasets. Traditional data-driven methods for equation discovery typically rely on single datasets, which can be limiting when observations are scarce. MCO-PDE addresses this by leveraging multiple datasets that describe the same physical system but originate from different initial or boundary conditions. The framework operates by first training individual neural surrogates for each distinct data source. Subsequently, a soft-competitive weighting mechanism is employed to dynamically assess the credibility of each dataset and synthesize a consensus global coefficient. This process, integrated with a genetic algorithm for structural search, allows MCO-PDE to simultaneously identify both the functional forms and parameters of the underlying physical laws. Demonstrations show that MCO-PDE can accurately recover canonical equations using as few as 50 observations per dataset across various cases. It effectively handles complex scenarios, including two- and three-dimensional domains with irregular boundaries and heterogeneous coefficients, and has successfully extracted meaningful physical laws from real-world wave-tank experiments.

Why it matters

This method significantly advances scientific machine learning by enabling the discovery of fundamental physical laws from fragmented or limited multi-source data, accelerating research and development in fields reliant on complex simulations and modeling.

How to implement this in your domain

  1. 1Apply MCO-PDE or similar multi-source learning techniques to discover governing equations in complex engineering systems with varied experimental data.
  2. 2Integrate this framework into scientific research pipelines to automate the derivation of physical models from observational data.
  3. 3Utilize the competitive optimization approach to improve model robustness when dealing with noisy or incomplete datasets from different sources.
  4. 4Explore using this method for inverse problems where underlying physical parameters need to be inferred from observed system behavior.

Who benefits

Scientific ResearchEngineeringMaterials ScienceClimate ModelingAerospace

Key takeaways

  • MCO-PDE discovers shared PDEs from multiple datasets using competitive optimization.
  • It trains independent neural surrogates and aggregates a consensus global coefficient.
  • The framework can accurately recover equations even with limited observations per dataset.
  • It handles complex domains and successfully extracts laws from real-world experiments.

Original post by Hao Xu, Siyu Lou, Yuntian Chen, Dongxiao Zhang

"arXiv:2606.30699v1 Announce Type: new Abstract: Discovering governing equations directly from observational data is a key step towards interpretable scientific machine learning. Current data-driven approaches typically operate on a single dataset, inherently limiting their perfor…"

View on X

Originally posted by Hao Xu, Siyu Lou, Yuntian Chen, Dongxiao Zhang on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses