Designing accurate emulators for scientific processes using calibration-driven deep models

Area of Science:

• Scientific Computing
• Machine Learning
• Data Science

Background:

• Predictive models, or emulators, accelerate scientific discovery by replacing slow numerical simulations or experiments.
• Machine learning is increasingly used to build data-driven emulators for complex scientific processes.
• Choosing appropriate loss functions is critical but often overlooked in emulator design.

Purpose of the Study:

• To address the limitations of symmetric loss functions (e.g., mean squared error, mean absolute error) in scientific emulators.
• To propose a novel deep learning approach for designing emulators that can recover complex noise structures without prior assumptions.
• To demonstrate the effectiveness of the proposed method across various use-cases, particularly in small-data scenarios.

Main Methods:

• Development of a novel deep learning approach named Learn-by-Calibrating.
• Utilizing interval calibration as the core mechanism within the deep learning framework.
• Testing the approach on a diverse suite of scientific use-cases.

Main Results:

• The Learn-by-Calibrating approach effectively recovers inherent noise structures in data.
• High-quality emulators were achieved, surpassing widely adopted loss function choices.
• The method demonstrates strong performance even in small-data regimes.

Conclusions:

• Learn-by-Calibrating offers a robust solution for designing scientific emulators capable of handling complex noise.
• The approach provides a valuable alternative to traditional loss functions, enhancing emulator accuracy and reliability.
• This method is particularly beneficial for applications with limited datasets or heterogeneous/asymmetric noise distributions.