Generative models and autonomous laboratories are collapsing the slowest part of science, the trial-and-error loop between an idea and a measured result. For India’s pharmaceutical, materials and instrument industries, the question is no longer whether this arrives, but who builds it.
From prediction to production
Two Nobel Prizes in 2024 announced, in effect, that machine learning had become a scientific instrument. The Physics prize honoured the foundations of neural networks; the Chemistry prize went to computational protein design and to the AI system that predicts protein structures from sequence. Within two years, the field has moved from predicting what molecules look like to designing new ones and, more consequentially, to letting machines decide which experiment to run next.
That last step is the real revolution. A generative model can propose thousands of candidate molecules, catalysts or protein variants in an afternoon. The bottleneck has always been the laboratory: synthesising those candidates, purifying them, measuring their properties and feeding the results back into the next design. This is the design–make–test–analyse (DMTA) cycle, and in conventional practice a single iteration can take weeks.
What a self-driving lab actually does
A self-driving laboratory (SDL) closes that loop without a human in the middle. Robotic arms and liquid handlers prepare reactions; automated analytical instruments such as LC–MS systems, plate readers and diffractometers characterise the products; and an optimisation algorithm, typically Bayesian optimisation or a reinforcement-learning agent, reads the data and chooses the next experiment. The system runs through the night, every night.
The evidence of acceleration is now concrete rather than promotional. In early 2026, Ginkgo Bioworks and OpenAI reported a study in which GPT-5 designed and directed more than 36,000 cell-free protein-synthesis experiments in Ginkgo’s cloud laboratory, driving down the cost of producing a benchmark fluorescent protein. Nature ran the work under a headline asking whether robot labs would replace biologists. At North Carolina State University, a materials SDL gathered roughly ten times more data than conventional methods and converged on optimal candidates in a fraction of the time. Lawrence Berkeley National Laboratory’s A-Lab has autonomously synthesised inorganic compounds by combining literature mining, density functional theory and robotic furnaces, identifying promising solid electrolytes in a few dozen autonomous trials.
The February 2026 SLAS conference in Boston confirmed the commercial turn: workcell vendors, scheduling-software firms and lab-informatics companies all launched AI agents for protocol conversion and experiment orchestration. The self-driving lab has moved from academic demonstration to procurement category.
Where India stands
India’s most visible entry is AILA, the Artificial Intelligence Laboratory Assistant developed at IIT Delhi with the National Chemical Laboratory. Reported in late 2025, it autonomously plans and executes experiments, and its developers cite a roughly thirtyfold increase in data throughput in a materials-science laboratory, with multistep organic synthesis as the next target. Indian-trained researchers are also prominent in the field abroad, building autonomous platforms for catalysis and semiconductor discovery at leading international groups.
The strategic case for India is stronger than its current footprint. India’s pharmaceutical and CDMO sector runs enormous volumes of process development, impurity profiling and formulation screening, precisely the repetitive, data-rich optimisation that SDLs do best. Process analytical technology (PAT) and digital twins, already adopted for regulatory quality-by-design filings, are the natural on-ramp: a plant that streams real-time spectroscopic data is halfway to a closed-loop optimiser.
There is also an industrial opportunity that is easy to miss. Self-driving labs are built from analytical instruments, liquid handlers, sample-preparation modules and control software. A country that aspires to manufacture laboratory instruments cannot treat autonomy as someone else’s integration problem. The interface standards, data formats and orchestration layers being settled now will decide whose hardware plugs in.
The honest limits
Claims that AI compresses drug or materials development “from decades to days” should be read carefully. What shrinks is the optimisation loop: finding a better catalyst composition, a higher-yielding reaction condition or a more stable protein variant. What does not shrink is everything downstream: toxicology, scale-up, stability testing, clinical trials and regulatory review. A molecule designed in a week still needs years to become a medicine.
Three further constraints deserve attention. First, data quality: an autonomous loop amplifies whatever systematic error sits in its instruments, so calibration, metrology and reference standards become more important, not less. Second, search-space bias: optimisers excel at exploring the region they are pointed at and can miss unexpected discoveries that a curious chemist would chase. Third, accountability: when an AI-directed laboratory produces a flawed or unsafe conclusion, responsibility must sit with identifiable humans, a point now being raised in automation-framework papers.
Sustainability is a quieter benefit. Autonomous platforms miniaturise reactions and extract more information per experiment, cutting solvent use and waste, a meaningful gain for an industry under pressure on green chemistry.
What Indian institutions should do now
Three moves would position India well. Research funders, including the Anusandhan National Research Foundation, could seed shared SDL facilities rather than isolated pilots, so that academic groups and start-ups rent autonomous experimentation the way they rent supercomputer time. Pharma and CDMO firms could treat their PAT data infrastructure as an AI asset and invest in clean, machine-readable experimental records. And instrument makers should design for automation from the start, with open control interfaces and standardised data output.
The laboratory of 2030 will not be empty of scientists. It will be a place where human judgement decides what questions are worth asking, and machines do the relentless work of answering them. The countries that build those machines, and the data that trains them, will set the pace of discovery.
