Fast Machine Learning for Science Conference 2026

Beyond Fast Inference: Toward Adaptive Scientific Systems

Not scheduled

20m

Topical session Birds-of-a-Feather Sessions

Speaker

Nhan Tran (Fermi National Accelerator Lab. (US))

Description

Scientific AI has made tremendous progress in accelerating inference, compressing models, and deploying machine learning at the edge. However, most scientific systems remain fundamentally open-loop: AI analyzes data or makes predictions, while scientists remain responsible for interpreting results, recalibrating instruments, updating models, and adjusting control strategies. The next frontier is enabling experiments to continuously perceive, predict, optimize, and adapt through tightly integrated, low-latency feedback loops.

This Birds-of-a-Feather session will explore a vision for adaptive scientific systems, where streaming measurements update state estimates, learned surrogate models provide rapid predictions, digital twins maintain an evolving representation of the experiment, optimization algorithms determine the next best action, and embedded controllers execute decisions in real time. Rather than treating AI as a standalone inference engine, this architecture embeds intelligence directly into the sensing and experimental control loop, enabling scientific instruments to continuously improve performance and respond autonomously to changing conditions.

This paradigm is broadly applicable to particle physics trigger systems, quantum control, accelerator tuning, fusion diagnostics and plasma control, autonomous laboratories, microscopy, and other experimental facilities where adaptive decision-making must occur under stringent latency constraints.

Rather than highlighting individual applications, this BOF will focus on the foundational challenges preventing adaptive scientific systems from becoming commonplace. What methodologies and workflows are needed to integrate streaming data, surrogate models, digital twins, optimization, and embedded control into a cohesive system? What software abstractions, interfaces, benchmarks, and open infrastructure are missing? How can adaptive systems continuously evolve while maintaining robustness, reproducibility, uncertainty awareness, and scientific trust?

By bringing together researchers across scientific disciplines, this session aims to identify shared methodological and infrastructure gaps, establish a common vocabulary for adaptive scientific systems, and begin defining a community roadmap for the next generation of intelligent scientific instrumentation.