How AI Became an Active Partner in Physics, Chemistry, and Biology

A robot named Adam, working inside a van-sized automated laboratory in the 2000s, produced what are considered the very first entirely automated scientific discoveries – modest findings about yeast biology that nonetheless marked a turning point. Two decades later, AI systems are not merely assisting scientists but actively participating in the discovery process itself: generating hypotheses, designing experiments, predicting molecular structures, and even controlling robotic labs that synthesize and test new materials around the clock.

The scale of this shift is staggering. Google DeepMind’s GNoME system has discovered 2.2 million new crystal structures, including 52,000 novel lithium-ion conductors, with external researchers already synthesizing 736 of those predictions in real labs. AlphaFold’s protein structure predictions earned Demis Hassabis a share of the 2024 Nobel Prize in Chemistry. Microsoft’s Discovery platform identified a new datacenter coolant prototype in just over one week – a process that traditionally takes months. These are not incremental improvements. They represent a fundamental restructuring of how science gets done across physics, chemistry, and biology.

From Tool to Collaborator: AI’s Evolving Role

AI’s place in the laboratory has evolved dramatically from simple data analysis to what researchers now describe as a collaborative “co-creator” capable of proposing hypotheses, designing experiments, and optimizing entire workflows. This transformation has been driven by advances in machine learning, generative models, and what are called agentic AI systems – programs that can reason, plan, and take action with minimal human oversight.

The shift gained decisive momentum with AlphaFold 2 in 2021, which solved a fifty-year grand challenge in biology by predicting protein structures from amino acid sequences with near-experimental accuracy. Its successor, AlphaFold 3, extends predictions to protein interactions with DNA, RNA, and small molecules, achieving at least a 50% improvement over existing methods. In materials science, AI foundation models pretrained on vast scientific datasets now support iterative loops combining prediction, physics-based validation, and robotics across disciplines.

Computational neuroscientist Sebastian Musslick of Osnabrück University captures the trajectory well: a year ago, he would have said there was a lot of hype. Now, he says, “there are actually real discoveries.”

Biology: Predicting Proteins and Designing Drugs

Biology has arguably seen the most visible AI breakthroughs. AlphaFold’s ability to predict protein structures – the three-dimensional shapes that determine how proteins function – opened the door to understanding drug-like molecules among an estimated 10^60 possibilities. Tools like BioEmu-1 now help decode protein dynamics, while novel enzymes designed entirely by AI have been validated at Berkeley Lab’s Advanced Light Source, bridging the gap between computational prediction and physical reality.

Drug discovery has been particularly transformed. Insilico Medicine used AI to identify a previously unknown protein involved in idiopathic pulmonary fibrosis – a devastating lung-scarring disease – and then designed a drug molecule to block it. That molecule, ISM001-055, became the first fully AI-designed drug to reach Phase II trials with positive results, developed in under 18 months at approximately 10% of traditional costs. BenevolentAI separately used AI to repurpose the drug baricitinib for COVID-19, receiving FDA authorization within a year.

At Caltech, Frances Arnold’s lab – already renowned for directed evolution work – now uses generative AI and neural networks to mutate gene sequences for novel enzymes. Her team barcodes experimental data specifically for AI training, potentially enabling what she describes as one-button enzyme design executed by robots. Meanwhile, the Morpheus neural network at Caltech predicts tumor alterations for immune therapy susceptibility by analyzing seqFISH data on DNA, mRNA, and proteins simultaneously, identifying specific gene expression tweaks such as upregulating two genes and downregulating one.

Chemistry: Physics Meets Generative AI

In chemistry, a “second wave” of AI-driven discovery is emerging – one that goes beyond generic machine learning models trained on historical data. The key differentiator is physics. Companies and research groups now combine computational physics, structure-based modeling, and generative AI to design novel molecules from scratch, even for targets with no prior experimental data.

Aqemia’s QEMI engine exemplifies this approach. It runs an iterative loop between generative AI and physics-based evaluation, using quantum-inspired physics to test and refine large numbers of candidate molecules in silico before any synthesis occurs. As Aqemia’s CEO Maximilien Levesque puts it: “One major misconception is that AI models trained purely on historical data can drive true drug invention – you don’t invent by remixing history.”

Materials chemistry has seen equally dramatic advances. Berkeley Lab’s A-Lab facility operates a tight loop between machine intelligence and automation: AI algorithms propose new compounds, and robots prepare and test them, drastically shortening the time to validate materials for batteries and electronics. The lab’s Autobot robotic system at the Molecular Foundry investigates new materials for applications ranging from energy storage to quantum computing. Microsoft’s AI screened more than 32 million candidates to discover a new material that could reduce lithium use in batteries by up to 70%.

At Caltech, the C-CAS center led by Hosea Nelson and Sarah Reisman uses AI to analyze roughly 70% of lab reactions to predict optimal synthesis pathways for any molecule, uncovering hidden reaction patterns that could substantially reduce pharmaceutical manufacturing costs.

Physics: Simulations, Fusion, and Weather

Physics applications of AI tend to focus on optimizing complex simulations and controlling systems that operate at scales where human intuition breaks down. At the National Energy Research Scientific Computing Center (NERSC), machine learning predicts particle behavior in fusion plasmas, with the potential to inform live control systems for future fusion reactors.

GNoME’s 2.2 million new crystal structures target conductors for batteries and quantum technology. In materials physics, Microsoft’s new AI model for Density Functional Theory helps solve a 60-year challenge by quickly and accurately simulating how electrons behave – a capability with applications spanning drugs, batteries, and green fertilizers.

Weather forecasting provides a striking demonstration of AI’s power. Neural network weather models now outperform traditional supercomputer simulations in 97% of scenarios and run 1,000 times faster. At Caltech, Anima Anandkumar’s team uses neural operators trained on 50,000 samples of historical weather data gathered at six-hour intervals over four decades. Her system runs on a single graphics processing unit yet matches the accuracy of models running on enormous supercomputers. When Hurricane Lee was brewing in September 2023, her test model correctly forecast landfall in Nova Scotia 10 days in advance, while standard European and U.S. models had the storm heading out to sea.

Theoretical physicist Alex Lupsasca discovered that OpenAI’s ChatGPT agent, running on GPT-5 pro, could independently find new symmetries in equations governing black hole event horizons – symmetries he had found through months of solo work. OpenAI confirmed the agent had not accessed his published paper. “I was like, oh my God, this is insane,” Lupsasca said. He subsequently moved to San Francisco to join OpenAI’s new science team.

Cross-Discipline Platforms and Automation

Some of the most consequential developments cut across traditional disciplinary boundaries. Microsoft Discovery, built with agentic AI, acts as a research teammate that automates tasks from forming hypotheses to running simulations and refining experiments. It recognizes patterns and connections across large datasets, helping scientists test ideas that would otherwise require months of manual work.

Laboratory automation represents another cross-cutting trend. Berkeley Lab’s infrastructure streams data from light sources, microscopes, and telescopes to the Perlmutter supercomputer, where it is processed within minutes. A web-based platform called Distiller enables real-time decision-making during experiments, allowing researchers to refine their approach while an experiment is still in progress.

Best Practices and the Risk of Junk Science

Speed without rigor is dangerous. Experts advocate what some call hybrid “systems of boxes” – stacking AI agents with verifiable tools such as physics simulations and curated databases to ensure accuracy. The core practices emerging from leading laboratories include:

• Iterative validation loops: Generative AI proposes candidates; physics-based evaluation refines them. QEMI runs continuous cycles. Berkeley Lab validates AI-designed enzymes at the Advanced Light Source.
• Automation integration: Real-time AI-robotics workflows enable 24/7 experimentation, reducing human repetition while maintaining consistency.
• Physics grounding: For chemistry and physics applications, structure-based models outperform generic ML when generating first-in-class molecules, especially for targets with no prior ligand data.
• Agentic workflows with guardrails: Reasoning and planning AI systems like Discovery run hypothesis-to-simulation pipelines, but human experts must remain in the loop to catch compounding errors.

The risk of AI-generated junk science is real. In 2025, the journals PLOS and Frontiers stopped accepting submissions based only on public health datasets because too many papers were AI-generated slop. NYU cognitive scientist Gary Marcus warns that large language models’ biggest scientific application so far may be “writing junk science.” At the first scientific meeting for AI-agent-led research in October 2025, human attendees noted that AI frequently made mistakes, and one team published a paper detailing why LLM-based agents are not yet ready to be autonomous scientists.

The Uneven Distribution Problem

For all its promise, AI’s benefits in science remain unevenly distributed. When researchers at under-resourced institutions attempted to apply methods similar to AlphaFold, they lacked the computational infrastructure, curated datasets, and interdisciplinary expertise required. The breakthroughs concentrate in domains with strong data infrastructure, well-coordinated research communities, and access to computational resources that most institutions cannot match.

Addressing this requires more than technical innovation. Shared platforms, open tools, and cross-disciplinary education are essential. Berkeley Lab’s approach of pioneering AI-enabled discovery platforms and sharing them across the research community offers one model. Free-access tools designed specifically for students and researchers are making enterprise-grade capabilities more broadly available. But the gap remains significant.

What Comes Next

The trajectory points toward AI embedded not just in data analysis but in experimental design, real-time instrument control, and operational decision-making across every stage of the scientific process. Emerging trends include self-supervised learning on unlabeled data, geometric deep learning for structured data like molecules, and symbolic regression for discovering equations directly from experimental observations.

AI can now uncover “latent knowledge” buried in past publications, predicting functional materials years before laboratory discovery confirms them. It predicts black hole spacetimes, simulates particle collisions, and identifies hydropower sites by finding patterns embedded in historical data that human researchers never noticed.

The technology is not replacing scientists. Frances Arnold notes that AI reveals mutational insights “not obvious to the human brain,” but her lab had to change how it collects and barcodes data to make AI useful. Caltech’s Tapio Schneider pretrains models on physics simulations before fine-tuning with satellite observations. The most productive path forward combines human creativity with machine capability – what Anandkumar calls AI “fundamentally transforming the whole scientific method” while remaining dependent on the quality of human questions and the rigor of experimental validation.

As OpenAI’s Kevin Weil puts it about AI-enabled discovery: “We’re still totally in the early days.” But the early days are already producing Nobel Prizes, millions of new material candidates, and drugs reaching clinical trials at a tenth of traditional costs. The scientific method itself is not changing – but the speed, scale, and scope at which it operates will never be the same.

Sources

• AI for Scientific Discovery is a Social Problem
• AI for Science: 5 Ways It’s Helping Solve Big Challenges
• Have We Entered a New Age of AI-Enabled Discovery?
• Generative AI for Science – Comprehensive Guide
• How Physics Is Shaping the Next Wave of Drug Discovery
• MIT Professional Education: AI for Scientific Discovery
• Caltech: How AI Advances Scientific Discovery
• Berkeley Lab: AI and Automation Speed Up Discovery
• Microsoft Research: AI Case Studies for Natural Science