Multi-agent system improves scientific design optimization with learning
ADAPT for Science: A Multi-Agent System for Self-Evolving Scientific Optimization
Artificial IntelligenceMultiagent Systems
Summary
Finding the best designs for scientific experiments can be very hard because testing each option takes a lot of time and resources. The authors created ADAPT for Science, a system where multiple agents work together to decide not only where to look for good designs but also how to search better over time. They tested their system on designing parts of gravitational wave detectors and showed it works much better than traditional and other automated methods. This system learns from past designs and its own search results to improve continuously.
What this means in practice
- •For engineering design teams: Improve design optimization for complex systems by using a self-evolving multi-agent framework that adapts from past attempts and search feedback.
- •For hardware development teams: Optimize hardware configurations under strict physical constraints more efficiently by integrating learned search strategies from historical design data.
Authors
Zhaoran Feng, Yuchen Ni, Jianan Liu, Lemao Liu, Dejing Dou, Kairui Feng
Abstract
Scientific discovery often requires identifying a small number of high-performing designs in vast, constrained search spaces. This is difficult when evaluations are costly, feasible solutions are sparse, parameters are strongly coupled, and physical constraints are strict. Recent AI for AI methods can automatically improve algorithms and optimization procedures. However, general purpose approaches are not designed around scientific priors, physical constraints, and limited evaluation budgets. We introduce ADAPT for Science, a multi-agent framework that treats scientific optimization as two coupled decisions: where to search and how to search. ADAPT uses historical designs to identify promising and diverse initial conditions, progressively concentrates computation on favorable search directions, and evolves the optimization procedure from execution feedback. We evaluate ADAPT on the design of gravitational wave detectors using a differentiable physics simulator and eight unseen detector topologies. Under matched scientific evaluation budgets, ADAPT achieves a mean loss of 0.204 +/- 0.101, reducing loss by 54.9% relative to the best classical optimizer and by 62.2% relative to the best general purpose agentic evolution baseline. ADAPT achieves the lowest loss on all unseen topologies among the evaluated baselines. ADAPT provides a step toward scientific optimization systems that continuously learn from prior experience and their own search outcomes.