How Bayesian optimization, transfer learning, and closed-loop workflows are reshaping molecule discovery by making search faster, smarter, and more targeted.
The sheer scale of the experimental design space has long been a bottleneck in molecular discovery. From small-molecule pharmaceuticals to advanced performance materials, researchers must navigate millions of potential combinations, yet all these efforts are constrained by limited experimental capacity.
Artificial intelligence (AI) is changing this game, shifting R&D away from resource-intensive brute-force screening toward informed navigation. Instead of designing experiments manually, scientists can now rely on easy-to-use AI-based tools that guide the search through this immense design space. These tools recommend the most informative next experiment based on results collected so far. Scientists then run the recommended experiment, feed the results back into the system, and repeat the loop until convergence.
By automating the heavy lifting of experimental design, these tools free scientists to focus on what matters most: interpretation, analysis, and creative problem-solving. The result is a more efficient, digitalized discovery process that turns every experimental outcome into compounding knowledge.
Artificial intelligence (AI) is changing this game, shifting R&D away from resource-intensive brute-force screening toward informed navigation. Instead of designing experiments manually, scientists can now rely on easy-to-use AI-based tools that guide the search through this immense design space. These tools recommend the most informative next experiment based on results collected so far. Scientists then run the recommended experiment, feed the results back into the system, and repeat the loop until convergence.
By automating the heavy lifting of experimental design, these tools free scientists to focus on what matters most: interpretation, analysis, and creative problem-solving. The result is a more efficient, digitalized discovery process that turns every experimental outcome into compounding knowledge.
Table of Contents
- Main Takeaway: AI as an Actionable Decision Engine
- From Edisonian Trial-and-Error to the Engine of Efficiency: Bayesian Optimization
- Redefining Reaction Generalization via Transfer Learning
- Optimizing Across Information Frontiers: Multi-Fidelity Bayesian Optimization
- Physical AI in the Lab: Self-Driving Labs®
- Transforming R&D from Linear to Exponential
Main Takeaway: AI as an Actionable Decision Engine
AI accelerates molecule discovery and modern drug design when it goes beyond prediction to serve as an active decision engine. Traditional machine learning models can predict properties of specified molecules, but still leave the burden of experimental design on the scientist. True value emerges when optimization algorithms, historical knowledge transfer, and closed-loop execution combine into a single framework that determines the next most informative step to take. This lets teams spend less time on resource-intensive trial-and-error and more time testing the parameters that drive high-value discovery. Every experiment then becomes part of an automated learning process, allowing research to compound sequentially instead of resetting from zero with each new campaign.
From Edisonian Trial-and-Error to the Engine of Efficiency: Bayesian Optimization
For decades, molecule discovery has followed an Edisonian approach: synthesize a candidate, measure its performance, adjust parameters, and repeat. This “make-then-measure” logic still works well when chemical spaces are small and problems are tightly constrained. But it struggles in higher-dimensional problems, where the number of possible molecules, reaction conditions, and formulations quickly outpaces what brute-force experimentation can realistically explore.
To break through these bottlenecks, AI-assisted discovery reframes the challenge as an optimization problem. Rather than searching exhaustively, researchers first define a rigorous objective function that quantifies the desired properties (e.g., maximizing yield and selectivity while minimizing raw material costs). The system then searches intelligently across the multi-dimensional design space to optimize these objectives. Bayesian Optimization (BO) provides a principled framework for this search, using a probabilistic model of the objective to decide which experiments to run next based on what is known versus what is merely guessed.
To break through these bottlenecks, AI-assisted discovery reframes the challenge as an optimization problem. Rather than searching exhaustively, researchers first define a rigorous objective function that quantifies the desired properties (e.g., maximizing yield and selectivity while minimizing raw material costs). The system then searches intelligently across the multi-dimensional design space to optimize these objectives. Bayesian Optimization (BO) provides a principled framework for this search, using a probabilistic model of the objective to decide which experiments to run next based on what is known versus what is merely guessed.
→
At its core, BO is designed for environments where every experiment is an investment of time, precursors, and analytical throughput: the exact reality of most chemistry workflows. While traditional machine learning (ML) requires massive datasets just to start making predictions, and classical Design of Experiments (DoE) forces researchers into a rigid, non-adaptive grid, BO thrives in these data-scarce scenarios. It operates in a sequential, data-efficient way: starting from a handful of initial runs, the algorithm treats every data point as high-leverage information and compresses the timeline needed to hit target performance thresholds.
The Mechanics Behind the Intelligent Search & Bayesian Optimization
To achieve this extreme data efficiency, the algorithm relies on a clear division of labor between two core components: a predictive model and a decision rule.
- The Probabilistic Surrogate Model (The Predictive Model): Typically driven by a Gaussian Process (GP), this functions as a dynamic digital twin of the multi-dimensional reaction landscape. Based on sparse initial runs, it estimates how the reaction will behave across untried conditions. Crucially, because it is probabilistic, it doesn’t just give a blind prediction; it quantifies its own uncertainty, mapping out where it has high confidence based on hard data and where it is merely guessing.
- The Acquisition Function (The Decision Rule): This component acts as the decision rule that scans the landscape to select the next experiment. Instead of a scientist manually weighing competing variables, the algorithm balances two core optimization criteria simultaneously:
- Exploration: Directing the next run into highly uncertain, unmapped territory to see if a superior catalyst backbone, solvent system, or formulation is hiding there.
- Exploitation: Steering toward known, high-performing regions to make small adjustments to the optimization variables and hit the peak performance target.
With each measurement, the algorithm updates its understanding of the space and moves toward optimal conditions step by step. This sequential design is what makes BO well-suited to experimental research, where data is expensive and scarce. Because its model is probabilistic, it is aware of where it is not confident, so even a handful of data points yields useful, honestly calibrated predictions, and each new experiment can be chosen where the algorithm will learn the most.
Other AI approaches typically lack this fit. For example, point-prediction ML models output a single best guess with no sense of their own reliability, so they need large datasets to become trustworthy and fail silently when data is scarce. For large language models (LLMs), they are trained on text in which real-world experimental data is barely represented, leaving them without grounding in the physical and chemical behavior of actual systems.
Other AI approaches typically lack this fit. For example, point-prediction ML models output a single best guess with no sense of their own reliability, so they need large datasets to become trustworthy and fail silently when data is scarce. For large language models (LLMs), they are trained on text in which real-world experimental data is barely represented, leaving them without grounding in the physical and chemical behavior of actual systems.
Case Study: Cutting Catalyst Costs Without Sacrificing Performance
The real-world power of this framework was demonstrated in a recent hydroformylation study. A process chemistry team faced a massive reaction space of up to 2.9 billion potential parameter combinations, with the critical objective of drastically reducing expensive rhodium catalyst loading without sacrificing conversion or selectivity.
Figure from graphical abstract Saudan et al. Bayesian Optimization for Resource-Efficient Hydroformylation. ACS Catalysis. 2025.
Atinary’s BO engine within SDLabs® AI platform, converged to robust, optimized catalyst conditions in only 88 sequential experiments. The platform reduced rhodium usage by up to 30x and its cost contribution by 97%, demonstrating that when intelligent algorithms are paired with domain expertise, even vast design spaces become tractable.
Learning From Every Campaign:
Transfer Learning Across Substrates
Discovering a molecule is not just identifying a promising structure; it means being able to make it. That requires a working reaction with the right conditions, and the same reaction rarely behaves the same way on every molecule. Conditions optimized for one starting material, called a substrate, often fail on the next, so chemists must re-tune them again and again, a problem known as “reaction generalization.
Transfer learning breaks this cycle. Rather than forcing one rigid recipe onto every substrate, it maps the underlying features of the reaction landscape and carries them forward. Knowledge from past (“source”) campaigns narrows the search space for each new (“target”) campaign, so the model can converge on the right conditions for an unfamiliar substrate with only a few experiments.
Caption: Figure adapted from graphical abstract Peng et al. Accelerating Reaction Generalization through Domain-Specific Transfer Learning. Chemrxiv. 2026.
A recent joint ChemRxiv publication between the lab of Dr. Alan Healy at NYU Abu Dhabi and Atinary. Using Atinary’s SDLabs, the collaboration provided a definitive demonstration of how transfer-learning-guided workflows can accelerate substrate-specific reaction optimization.
Instead of relying on large, literature-derived datasets that carry their own biases, the workflow operates via an efficient three-step framework:
- Explore: The team navigated the initial reaction landscape using BO in SDLabs, building a compact, domain-specific dataset of just 120 experiments from an expert-guided selection of the reaction space.
- Adapt: Using Atinary’s proprietary transfer learning algorithm, SeMOpt (Semantic Memory Optimization), this curated data was applied as “semantic prior knowledge” to entirely new, unseen substrates.
- Optimize: By balancing this semantic knowledge with targeted exploration-exploitation, the model converged on optimal, substrate-specific conditions much faster, in four experiments or fewer.
By pairing human expertise with this transfer learning architecture, R&D teams can achieve genuine operational generalization. The approach replaces blind, repetitive trial-and-error with a fast, smart discovery loop that adapts quickly to new substrates.
Physical AI in the Lab: Self-Driving Labs®
Self-Driving Labs® are where these individual data-driven pillars converge. Within a fully autonomous loop, optimization models propose experiments, robotic systems execute them, analytical tools characterize the output, and the resulting data feeds straight back into the model. This seamless integration of software, hardware, and data, turns AI from a passive recommendation engine into an active engine of continuous discovery, shifting the pace of R&D.
A definitive look at this architecture in practice is our state-of-the-art Atinary Lab in Boston. Operating as a Scientific Discovery Factory, the facility features two autonomous platforms that continuously Design, Make, Test, Analyze and Learn (DMTA+L) from real-world experiments around the clock.
A definitive look at this architecture in practice is our state-of-the-art Atinary Lab in Boston. Operating as a Scientific Discovery Factory, the facility features two autonomous platforms that continuously Design, Make, Test, Analyze and Learn (DMTA+L) from real-world experiments around the clock.
Atinary’s Self-Driving Labs® integrate our cloud-native, code-free AI platform (SDLabs) with world-class automation and instrumentation from an ecosystem of industry-leading partners:
- Robotics & Automation: ABB Robotics and Chemspeed Technologies provide configurable automation systems that handle the physical preparation and execution of complex reactions with high precision.
- Precision Measurement & Synthesis: Mettler-Toledo delivers the laboratory instrumentation required for high-fidelity physical operations and sample handling.
- Advanced Analytics: Agilent (Open Bed Sample Fraction Collector and Liquid Chromatography and Mass Spectrometer) and Bruker (benchtop NMR) provide high-throughput tools that characterize reaction outputs in real time.
- Cloud Infrastructure: Hosted on Amazon Web Services (AWS), Atinary’s SDLabs AI platform provides the orchestration layer for a secure, scalable architecture, ensuring seamless machine-to-machine communication at scale. It combines cloud-native compute and secure data environments with modern generative AI (such as Amazon Bedrock) to give bench scientists an intuitive conversational interface to the system.
By focusing initially on small-molecule synthesis and catalysis, specifically complex couplings like the Buchwald-Hartwig and Suzuki reactions, the Atinary Lab in Boston serves as a living blueprint for the future of R&D. It bridges digital simulation and physical reality, running continuous DMTA+L cycles that generate more high-quality, ML-ready data in a single week than a traditional workflow produces in years.
Where This Leads: A Unified Discovery Loop
One idea runs through all of these approaches: AI pays off in the lab when every experiment sharpens the next. Bayesian optimization decides which experiment is worth running, transfer learning carries forward what past campaigns have already taught, and autonomous execution keeps the loop running around the clock. Together they turn discovery into a system that compounds, each result making the next one faster and better informed.
This is what Atinary’s SDLabs® platform and Self-Driving Lab in Boston are built to deliver: freeing scientists from the mechanics of experimentation to focus on interpretation, analysis, and deciding what to pursue next. The result is a repeatable, sustainable path to the right molecules and performance materials, a competitive advantage for the teams that adopt it, and a faster route to the breakthroughs that pressing global challenges demand, in human health, sustainable food, climate, and clean energy.
This is what Atinary’s SDLabs® platform and Self-Driving Lab in Boston are built to deliver: freeing scientists from the mechanics of experimentation to focus on interpretation, analysis, and deciding what to pursue next. The result is a repeatable, sustainable path to the right molecules and performance materials, a competitive advantage for the teams that adopt it, and a faster route to the breakthroughs that pressing global challenges demand, in human health, sustainable food, climate, and clean energy.
Author Contributions
Edlyn Wu, PhD (Scientific Solutions Associate)
Mohammed Azzouzi, PhD (Applications Engineer)
Mahrokh Boroujeni (AI Research and Innovation Team Lead)
Edlyn Wu, PhD (Scientific Solutions Associate)
Mohammed Azzouzi, PhD (Applications Engineer)
Mahrokh Boroujeni (AI Research and Innovation Team Lead)