How Computational Chemistry is redefining Self-Driving Labs

Last week, Nextmol participated in the Intelligent Labs for Science event, organized by IMDEA Energy, a gathering focused on AI-driven discovery, autonomous laboratories, and the future of science. It was the kind of event that makes you pause and realize: the way we do chemistry is fundamentally changing. And the change is happening faster than most of the industry appreciates.

One of the central themes throughout the event was the concept of the self-driving laboratory, a system capable of autonomously planning, running, and learning from experiments in a continuous loop, without waiting for a human to interpret results before taking the next step. It’s a compelling vision. But there’s a version of it that’s already here, already industrially proven, and routinely overlooked in the conversation: the computational laboratory.

What makes a lab “self-driving“? 

The term “self-driving lab” typically evokes robotic arms, automated liquid handlers, and AI systems that schedule physical experiments around the clock. This version is exciting and increasingly real. But striping the concept to its essence, a self-driving lab is defined by three capabilities:

1. Plan: autonomously design the next experiment based on goals, constraints, and prior results
2. Run: execute that experiment without human intervention
3. Learn: extract insight from the output, update the model, and feed it back into the next plan

This Plan–Run–Learn loop is the architecture that transforms a laboratory from a place where humans conduct experiments into a system that generates knowledge autonomously. And computational labs, powered by molecular modeling and AI, operate on exactly this architecture.

The Computational Lab as an autonomous discovery engine

At Nextmol, we believe that computational laboratories, when powered by molecular modeling, enable a form of high-throughput screening that is complementary to physical labs, both in speed and scope.

Consider what a molecular dynamics (MD) simulation actually does. Given the structure of a molecule or formulation, it computes the forces between every atom and propagates the system forward in time, generating an atomic-resolution picture of how your material behaves. Properties like glass transition temperature, viscosity, solubility, stress-strain response, or interfacial tension can be predicted computationally, before a single gram of material is synthesized.

This is already a powerful capability on its own. But the transformation happens when you combine MD simulation with AI-driven orchestration:

– The “Plan” step becomes an AI agent that selects the next most informative molecular candidates to simulate, guided by Active Learning or Bayesian optimization, prioritizing experiments that will reduce uncertainty the most.
– The “Run” step is automated simulation pipelines executing across high-performance computing infrastructure, running hundreds of molecular systems in parallel.
– The “Learn” step is a machine learning model that ingests the simulation results, updates its predictive capability, and narrows the search toward the most promising region of chemical space.

The result is a closed-loop system that can explore vast regions of the chemical space — including molecules that have never been synthesized — building consistent, high-quality datasets at a pace no physical lab could sustain.

The data problem that physical labs can’t solve

There’s a less-discussed reason why the computational lab is essential in the self-driving laboratory ecosystem: data quality and coverage.

Machine learning models for chemistry are only as good as the data they’re trained on. Physical experimental data, while invaluable, comes with inherent limitations: it’s expensive to generate, prone to batch-to-batch variability, often scattered across disconnected internal databases, and critically, it only covers molecules that someone has already synthesized and decided to test.

This last point matters enormously. The chemical space of drug-like molecules alone is estimated at 10^60 compounds. The fraction that has ever been tested experimentally is vanishingly small. Any AI model trained exclusively on experimental data is, by definition, operating on a tiny and biased sample of what’s possible.

Computational simulation changes this equation. It can generate physically grounded property predictions for molecules that don’t yet exist in any laboratory, systematically, at scale, and with controlled uncertainty. This is what we mean when we talk about extensive sampling of the chemical space: not just running more experiments on known molecules, but exploring novel chemistries that lie beyond the experimental frontier.

When physical and computational labs work together — the computational system generating the training data and filtering the candidates, the physical lab validating the most promising ones — you get a genuinely powerful self-driving engine, accelerating discovery by orders of magnitude.

Why this matters for Specialty Chemicals

In the specialty chemicals industry, performance differentiation, sustainable formulation, and speed to market are the competitive axes that matter. All three are being reshaped by AI.

BCG’s recent research on AI in chemicals⁽¹⁾ estimates that full-potential AI transformation could unlock more than 3 percentage points of EBITDA margin for specialty chemical players, with R&D as one of the highest-value levers. The mechanism is straightforward: AI compresses the innovation cycle. Instead of iterating through dozens of physical formulations to find the one that works, teams can enter the lab with a computationally pre-filtered shortlist of high-confidence candidates. Fewer dead ends. Faster development. Lower cost of failure.

The self-driving computational lab is the infrastructure that makes this possible. It’s not a future aspiration. It’s operational today, in projects spanning polymer design, surfactant optimization, lubricant formulation, and sustainable materials development.

Reference:
(1) BCG Executive Perspectives “AI as the new performance divide”, March 2026