Princeton University

Using neural ordinary differential equations (NODEs) optimized simultaneously across 290 mass-ratio time series of ice crystals grown in a levitation diffusion chamber, the study learned the unknown functional form of the transfer coefficient G in the depositional ice growth model, then used symbolic regression (PySR) to derive a closed-form expression; the weakly-constrained NODE model performed best on 138 of 290 experiments (MSE loss 16495, versus 41396 for no surface kinetics and 40751 for Nelson and Baker 1996), yielding G = a0·Gc^a1·[a2 + a3/m]^-1 + a4 with an added ice-crystal-mass term, and this expression predicted ice supersaturation and ice water content better than the capacitance growth model and Nelson and Baker 1996 in independent AIDA Aerosol and Cloud Chamber experiments.

Using neural ordinary differential equations (NODEs) optimized simultaneously across 290 mass-ratio time series of ice crystals grown in a levitation diffusion chamber, the study learned the unknown functional form of the transfer coefficient G in the depositional ice growth model, then used symbolic regression (PySR) to derive a closed-form expression; the weakly-constrained NODE model performed best on 138 of 290 experiments (MSE loss 16495, versus 41396 for no surface kinetics and 40751 for Nelson and Baker 1996), yielding G = a0·Gc^a1·[a2 + a3/m]^-1 + a4 with an added ice-crystal-mass term, and this expression predicted ice supersaturation and ice water content better than the capacitance growth model and Nelson and Baker 1996 in independent AIDA Aerosol and Cloud Chamber experiments.

This work introduces a modular, large language model (LLM)-powered agent pipeline that automatically extracts and structures composition–process–property attributes of concrete materials from tables and text in scientific publications, achieving F1 scores up to 0.98 across 17 open and proprietary models and, within about one hour, extracting over 10,000 records from 278 papers screened from more than 27,000 publications; after postprocessing this yields the largest open laboratory database for blended cement concrete with nearly 9,000 high-quality records and over 100 attributes, and machine learning analyses indicate that large, diverse, information-rich datasets improve both in-distribution accuracy and out-of-distribution generalization to unseen materials systems.

This work introduces a modular, large language model (LLM)-powered agent pipeline that automatically extracts and structures composition–process–property attributes of concrete materials from tables and text in scientific publications, achieving F1 scores up to 0.98 across 17 open and proprietary models and, within about one hour, extracting over 10,000 records from 278 papers screened from more than 27,000 publications; after postprocessing this yields the largest open laboratory database for blended cement concrete with nearly 9,000 high-quality records and over 100 attributes, and machine learning analyses indicate that large, diverse, information-rich datasets improve both in-distribution accuracy and out-of-distribution generalization to unseen materials systems.