Collaborative research: caig: advancing long-term temperature reconstruction using physics-informed neural networks through lake model emulation

The National Science Foundation Office of Integrative Activities awarded the University of Montana $276,152 on August 15, 2026, under the Geosciences program (CFDA 47.050) to develop physics-informed machine-learning emulators of lake thermal process models for global temperature reconstruction from sediment cores.

The project replaces computationally intensive first-principles physical lake models with machine-learning surrogates trained to emulate those models, enabling dramatic speedup of lake simulations. The team will build a physics-informed neural network emulator of a widely used lake process model and develop a generative machine-learning pipeline to downscale model outputs to the spatial and temporal scales required by lake simulations. The emulator will be coupled with Bayesian calibration, uncertainty quantification, and inverse-modeling methods to construct a uniform global temperature dataset from the full archive of lake sediment cores collected worldwide. The resulting long-term local temperature records will strengthen the scientific foundation for seasonal and long-term prediction, supporting more informed decisions across weather-sensitive sectors of the U.S. economy.

Work is performed in Missoula, Montana, with a period of performance through July 31, 2029. The assistance type is a Project Grant.

Abstract

This project harnesses the power of artificial intelligence (AI) to turn environmental observations from multiple sites into uniform data products that can be used to conduct scientific research and address problems of national need. The specific observations here are sediment cores collected from lakes that can be analyzed to infer local temperature variations over thousands of years. The traditional method for inferring temperature from sediment cores relies on a first-principles physical model to simulate the lake, a reasonable strategy for a single lake but impractical as a means to consistently derive temperature records from all lake sediment cores collected around the world. The solution pursued here replaces the physical lake model with a machine-learning (ML) derived surrogate trained to emulate the lake model. The emulator allows a dramatic speedup of lake simulations, and combined with other ML-based strategies allows the construction of a uniform global temperature dataset based on the full archive of lake cores. This project will result in improved understanding of long-term variability in local temperatures, which will strengthen the scientific foundation for seasonal and long term prediction, enabling more informed decisions across weather-sensitive sectors of the U.S. Economy. The project will develop a general framework for creating physics-informed ML emulators of geoscience process models and apply IT to lake thermal dynamics. The team will build a globally generalizable emulator of a widely used lake process model using physics-informed neural networks, and will develop a generative ML pipeline to downscale model outputs to the spatial and temporal scales required by lake simulations. The emulator will be coupled with bayesian calibration, uncertainty quantification, and inverse-modeling methods to investigate whether seasonal biases in proxy records contribute to the long-standing mismatch between modeled and reconstructed temperatures during the past 11,700 years. By enabling computationally efficient simulations for more than 100 lakes, the project will make IT possible to evaluate seasonal proxy biases at an unprecedented scale. The research will also systematically assess the accuracy and performance tradeoffs between emulators and traditional physical models, producing guidance for when emulators can effectively supplement or replace computationally intensive physical models. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the foundation's intellectual merit and broader impacts review criteria.- subawards are not planned for this award.