Researchers from the Centre for Advanced Laser Applications have developed a novel approach that combines multi-objective and multi-fidelity optimization, which could significantly speed up the optimization of complex systems like laser-plasma accelerators. The work, recently published in IOP Machine Learning: Science and Technology, introduces the innovative use of a trust metric to facilitate the joint optimization of multiple objectives and data sources.

The new "Trust-based Multi-Objective Multi-Fidelity" (Trust-MOMF) optimization method modifies the standard multi-objective Bayesian optimization approach by incorporating the trust gain per evaluation cost as an additional objective. This allows the algorithm to simultaneously optimize multiple objectives while leveraging low-fidelity, computationally cheap approximations to efficiently establish the Pareto set of optimal solutions.

When applied to particle-in-cell simulations of laser-plasma acceleration, a highly complex and computationally expensive problem, the Trust-MOMF approach yielded optimization results in a fraction of the time compared to standard methods. "Our method can reduce the computational cost of multi-objective optimization by an order of magnitude," said Dr. Andreas Döpp, who lead the study. "This will enable us to explore a much wider parameter space in simulations and experiments."

By providing an efficient way to optimize systems with multiple, potentially competing objectives, the Trust-MOMF method paves the way for accelerating both computational and experimental research, not just in laser-plasma physics but across various scientific computing domains. The team is already applying this approach to optimize laser-plasma acceleration experiments at the ATLAS-3000 laser facility.

Original publication:

Leveraging trust for joint multi-objective and multifidelity optimization
F. Irshad, S. Karsch, A. Döpp
Machine Learning: Science and Technology 5, 015056 (2024)