Retraining on fresh simulations each step keeps the method accurate where fixed surrogates fail.

Scientists run computer models to mimic reality, but they need to know which input settings produce real-world observations. Finding this set of inputs is called source distribution estimationsource distribution estimationFinding a distribution of parameters that, when simulated, reproduces observed data.. Existing tools train a proxy model once using a fixed guess of settings. If the true settings fall outside that guess, the proxy fails. To solve this, researchers designed a repeating two-step method. First, it runs fresh simulations using its current best guess of the settings to train an updated helper model. Second, it averages what that helper model learned across the real observations to improve the guessed settings. Testing on three benchmark tasks, the new strategy beat existing tools, especially when tested on a classic model called Lotka--Volterra.

Why it matters

Scientists using computer models to study complex systems, like ecosystems or climate, need to infer hidden parameters from data. If their methods rely on an inaccurate approximation, they might draw wrong conclusions. This approach improves accuracy, leading to more reliable scientific insights.

Questions answered