Source-linked AI summary
The convergent laboratory: when AI reasoning, autonomous experiments, high performance and quantum computing reshape chemistry
Eliu Huerta, Xiaoyun Wang, Geetika Gupta, Edward H. Sargent, Cameron J. Owen, Victor Fung, Abhishek Mitra, Austin Cheng, Emma Bouchard, Shams Mehdi
TL;DR
Chemical discovery is increasingly constrained by the need to connect reasoning, computation, experimentation, and quantum calculations across modalities. This Comment describes how open scientific AI, self-driving laboratories, MLIPs, fault-tolerant quantum computing, and agentic frameworks can be integrated into a convergent discovery stack, with reported examples spanning accelerated catalyst discovery, reactive modeling, and reduced quantum resource estimates. Its central conclusion is that the components of this convergent laboratory already exist, while connecting them remains a formidable but tractable engineering challenge.
Problem
Chemical property prediction requires approximations, truncations, or finite resources, while disparate discovery modes lack a native interface for easy interaction.
Method
The Comment synthesizes scientific-reasoning AI, self-driving laboratories, MLIPs, fault-tolerant quantum computing, and agentic frameworks as an integrated discovery stack.
Results
The convergent approach includes catalyst discovery shortened from an estimated 60 weeks to 5 weeks, reactive modeling within 0.5 kcal mol−1 across 650 reactions, and a 4.3-fold reduction in estimated FeMoco Toffoli cost.
Takeaways & Limitations
The components of the convergent laboratory exist today, and connecting them is a formidable but tractable engineering challenge.
Abstract
from arXiv · showhide
This Comment emerges from TPC26 (https://tpc26.org), a conference convening leaders from academia, national laboratories, and industry who are reshaping materials science discovery. The meeting explored how AI, autonomous agents, self-driving labs, higher performance and quantum computing converge to amplify their individual impact on materials science discovery. The perspectives here reflect the firsthand experiences of researchers at these frontiers and capture the essence of this global endeavor. As AI-driven reasoning, autonomous agentic frameworks, self-driving laboratories, and fault-tolerant quantum processors mature simultaneously, we offer this Comment as a reference at what we believe is a tipping point of transformative advances and productive disruption in the chemical sciences.
Introduction
Chemistry is entering a convergence of scientific reasoning, autonomous experimentation, learned interatomic models, quantum computing, and agentic coordination. Together, these feedback loops can expand exploration beyond established chemical intuition.
- Introduction: Five technological currents—scientific-reasoning LLMs, self-driving laboratories, MLIPs, FTQCs, and agentic frameworks—are converging to reshape chemistry’s theory–computation–experiment interplay.The agentic layer connects otherwise disparate methods and technologies.
- Introduction: Accumulated domain knowledge enables exploration of compositional and structural spaces beyond established chemical intuition.This can accelerate access to materials and molecules that conventional approaches would not prioritize.
AI that reasons about chemistry
Scientific LLMs are advancing from text generation toward open, multimodal, and agentic systems that can be improved within reproducible research environments. Their supporting infrastructure enables domain-specific deployment and feedback-driven refinement.
- AI that reasons about chemistry: Nemotron models extend scientific LLMs toward reasoning over multimodal documents and integration into domain-specific research pipelines.The Omni model handles text, figures, and tables, while open models support local deployment, fine-tuning, transparency, reproducibility, and data privacy.
- AI that reasons about chemistry: NeMo-RL provides scalable reinforcement-learning infrastructure supporting SFT, DPO, GRPO, and DAPO across clusters from one GPU to hundreds of accelerators.This infrastructure allows reasoning models to improve using feedback from simulated or real scientific environments.
- AI that reasons about chemistry: Open-weight models and open-source reinforcement-learning infrastructure enable reproducible scientific AI research and community-driven improvement.These properties are difficult to achieve with proprietary closed APIs.
Closing the loop: self-driving laboratories
Self-driving laboratories have become operational systems that combine robotic experimentation, machine learning, and human chemical expertise. An editable glass-box workflow substantially shortened catalyst discovery in a CO2 electroreduction demonstration.
- Closing the loop: self-driving laboratories: A suite of seven self-driving laboratories spans inorganic materials, organic small molecules, medicinal chemistry, polymers, formulations, organ-mimicry systems, and scale-up.The systems demonstrate autonomous experimentation across varied chemical domains.
- Closing the loop: self-driving laboratories: 30 catalysts were tested in two hours using robotic synthesis, parallelized membrane-electrode-assembly testing, and a glass-box classifier.The classifier used categorical composition features and thermodynamic descriptors from AdsorbML.
- Closing the loop: self-driving laboratories: Human experts edited classifier shape functions to inject chemical intuition about adsorption intermediates such as *CO and *CHO.The GAM Changer interface made the model’s learning loop directly editable by scientists.
- Closing the loop: self-driving laboratories: 5 weeks replaced an estimated 60 weeks of conventional experimentation to discover a record-performing propylene-selective catalyst.The reported reduction followed integration of automation, machine learning, and human domain expertise.
Machine-learned potentials: bridging scales
Machine-learned interatomic potentials bridge fast force-field calculations and more accurate quantum chemistry by embedding physical structure into lightweight models. Task-specific variants extend this approach to reactive chemistry and industrial catalytic cycles.
- Machine-learned potentials: bridging scales: AIMNet2 provides hybrid-DFT-level accuracy for common non-metal and halogen elements while predicting energies, forces, and partial charges on a single GPU in seconds.It scales near-linearly with system size and supports high-throughput conformer screening that is prohibitive with DFT.
- Machine-learned potentials: bridging scales: AIMNet2 embeds charge equilibration, Coulomb interactions, and dispersion corrections rather than relying on a pure black-box architecture.This design keeps the model lightweight and fast while maintaining competitive accuracy.
- Machine-learned potentials: bridging scales: AIMNet2-Rxn achieves accuracy within 0.5 kcal mol−1 on Diels–Alder thermochemistry across 650 reactions while distinguishing endo versus exo and regioselectivity.The task-specific model addresses failures of standard MLIPs on bond-breaking and bond-forming pathways.
- Machine-learned potentials: bridging scales: AIMNet2-Pd accurately characterizes geometries across the Suzuki–Miyaura catalytic cycle on a CPU in minutes rather than requiring more expensive calculations.This fine-tuned model extends AIMNet2 to palladium-catalysed cross-coupling reactions relevant in industry.
Quantum computing: the high-fidelity anchor
Fault-tolerant quantum computing is presented as a route toward high-fidelity electronic-structure calculations, approaching exact finite-basis solutions with systematically controllable error. Progress in Hamiltonian representations and algorithms can reduce estimated resource requirements, although these results concern future hardware and active-space choices still require attention.
- High-fidelity electronic structure: Fault-tolerant quantum algorithms could approach full-configuration-interaction-quality energies with resource requirements scaling polynomially with system size and inverse target precision.High fidelity is defined relative to the exact solution of a chosen finite-basis electronic Hamiltonian.
- Resource reduction: 4.3× lower estimated Toffoli cost was achieved for the larger FeMoco active space with DFTHC relative to BLISS–THC.The result illustrates how Hamiltonian representations and quantum algorithms can reduce fault-tolerant quantum-chemistry resource estimates.
- Scope: These quantum-chemistry results are resource estimates for future fault-tolerant hardware, not calculations performed on present-day quantum processors.
- Chemical quantities and workflow: Beyond ground-state energies, fault-tolerant algorithms have been developed for molecular observables, expectation values, and real-time chemical dynamics.Active-space selection can strongly affect calculation quality and cost and may require expert judgment and iterative testing.
Autonomous Agentic Frameworks
Agentic frameworks address the lack of a native interface among disparate scientific operations by connecting reasoning, computational discovery, laboratory experimentation, analysis, and hypothesis refinement. They also support reproducible workflows and cross-domain data provenance.
- Framework integration: Agentic frameworks connect autonomous in silico discovery with experimental exploration within a single scientific framework.The framework described at Lila Sciences uses a central reasoning model across the scientific method.
- Framework integration: Lila’s agents span hypothesis generation, autonomous in silico discovery, wet and dry laboratory experimentation, analysis, and hypothesis refinement.
- Reproducibility: Agentic frameworks provide robust, reproducible workflows while handling data provenance across different domains.
- Applications: The framework has been used for discoveries in life and physical sciences, including mRNA therapies and catalysts for hydrogen production.
Convergence: the integrated discovery stack
The proposed discovery stack combines reasoning agents, agentic coordination, machine-learned interatomic potentials, robotic testing, and quantum-computing pathways. Its workflow assigns candidates to computational screening, direct experiments, or quantum treatment according to their electronic character.
- Reasoning and proposal: An LLM-based reasoning agent proposes candidate compositions and reaction conditions from scientific literature using reinforcement learning on chemistry environments.
- Computational screening: An agentic framework directs an MLIP to screen proposed candidates computationally before robotic testing.
- Candidate routing: Electronically straightforward candidates are flagged for direct robotic testing, while strongly correlated systems are routed to quantum computing.
Outlook
The convergent laboratory is framed as an emerging model in which AI agents reason, robots act, neural networks approximate, and quantum processors anchor chemical discovery. The central engineering challenge is integrating these existing capabilities while preserving openness, reproducibility, and human scientific judgment.
- Integration challenge: Chemical discovery is increasingly limited by integrating knowledge across text, spectra, simulations, and quantum measurements rather than by generating data alone.
- Convergent laboratory: The convergent laboratory combines AI reasoning, robotic action, neural-network approximation, and quantum processing in one discovery model.The passage presents its components as existing today.
- Implementation: Connecting these capabilities is described as a formidable but tractable engineering challenge already demonstrating power through autonomous frameworks.
- Research practice: The proposed integration should keep tools open and accessible, science reproducible, and human scientists responsible for asking the right questions.
Declarations
The article acknowledges NSF and Argonne National Laboratory LDRD support and identifies E.H. as writing lead, with all authors contributing to writing and review.
- NSF grants OAC-2514142 and OAC-2209892 supported E.H.
- Argonne National Laboratory provided LDRD funding under U.S. Department of Energy Contract No. DE-AC02-06CH11357.
- E.H. led the article’s writing, while all authors contributed to writing and review.