Scientific Foundation Models in the Physical World
Over the past two years, the SciFM conference has traced the rapid evolution of scientific AI: from the emergence of scientific foundation models that unify representation learning across domains, to the rise of agentic systems capable of reasoning, planning, and executing multi-step scientific workflows. These milestones have transformed how we approach discovery. Yet, a new frontier is emerging, one that shifts AI from observation and inference toward interaction with, and embodiment in, the physical world.
As AI models grow more multimodal, context-aware, and capable of grounded reasoning, the core scientific question is changing: How do we design, train, and deploy AI systems that understand, predict, and shape physical processes—from atoms to organisms, from materials to ecosystems, from molecules to engineered infrastructure?
Why This Theme Now?
-
Convergence of Models, Sensors, and Actuators
Scientific foundation models are no longer limited to text, code, or static datasets. They now integrate:-
simulation trajectories
-
molecular and materials structures
-
multimodal microscopy, spectrometry, and imaging
-
robotic execution logs
-
instrument control interfaces
-
real-time sensor streaming
-
-
This integration allows AI systems to perceive the physical world directly and close the loop between model → action → measurement → model update.
-
Rise of Scientific Agents With Physical Interfaces
Agents can now autonomously propose hypotheses, design experiments, drive lab instrumentation, and operate autonomous laboratories and field platforms.
The next step is physically grounded scientific autonomy, where agents interact with materials, biological systems, and engineered environments—bringing together AI and real-world physical dynamics. -
Transformational Impact Across Scientific Domains
As AI learns from real physical systems rather than static datasets, its capacity to accelerate discovery expands dramatically.
We envision SciFM 2026 as the premier venue to explore the intersection of scientific foundation models, embodied intelligence, and physical systems. Key topics include:
1. AI for Materials, Chemistry, and Manufacturing
-
Foundation models trained on quantum simulations, molecular dynamics, synthesis logs, and spectroscopy data.
-
Agents that design advanced materials, navigate synthesis parameter spaces, and drive autonomous synthesis robots and reactors.
-
Closed-loop manufacturing systems where AI monitors, controls, and optimizes in real time.
Exemplar directions:
autonomous materials discovery, catalysis optimization, battery and semiconductor design, process-condition agents for additive manufacturing.
2. AI for Biological and Living Systems
-
Multimodal biological FMs integrating genomics, proteomics, metabolomics, imaging, and phenotype data.
-
AI-guided design of experiments using robotics, microfluidics, and self-driving laboratories.
-
Learning physical constraints in cells, tissues, and microbial communities.
Exemplar directions:
LLMs for biology, protein engineering, metabolic engineering, drug design, autonomous experiments with organoids or microbes, field robotics for environmental sampling.
3. AI for Climate, Energy, and the Environment
-
FMs trained on climate simulations, satellite imagery, sensor networks, and geophysical datasets.
-
Earth-system “digital twins” that couple simulation and streaming observation.
-
Embodied agents for environmental monitoring, critical minerals exploration, carbon sequestration, and ecosystem restoration.
Exemplar directions:
carbon capture optimization, wildfire prediction, weather-simulation + observation fusion, renewable energy system optimization, autonomous geological sampling.
4. AI for Autonomous Labs, Facilities, and Infrastructure
-
Integration of foundation models with the DOE user facilities:
APS, ALS, LCLS, SNS, EMSL, JGI, ARM, and others. -
Agents that coordinate data acquisition, parameter steering, and resource scheduling.
-
Unifying scientific knowledge across facilities to support a nation-scale agentic laboratory network.
Exemplar directions:
AI-augmented beamlines, nanoparticle growth control, multi-modal imaging pipelines, robotic manipulation at facilities, multi-agent scheduling for HPC + labs.
5. AI + Robotics: Embodied Scientific Intelligence
The physical world requires embodiment. Scientific robots—from fixed platforms in wet labs to mobile field systems to humanoid lab assistants—are beginning to coordinate with agents and foundation models.
Key challenges:
-
calibration and control grounded in physics
-
realtime sensor-model fusion
-
bridging “symbolic reasoning” and low-level control
-
safe execution in dynamic lab and field environments
Exemplar directions:
humanoid robotics in labs, robotic crystallography, micro-manipulation, environmental field robots, closed-loop experiment orchestration.
6. Theoretical Foundations: Physics-Informed AI
AI systems must respect physical constraints. SciFM 2025 will highlight work on:
-
physics-informed neural operators
-
alignment of AI reasoning with physical laws
-
guarantees for stability, safety, and uncertainty
-
foundation models trained on simulation landscapes and PDEs
-
hybrid AI-simulation co-design frameworks
This connects foundational theory with domain practice.
SPEAKERS & PANELISTS
The speakers for the SCiFM26 Conference will be announced soon.
