A practical playbook for moving from trial-and-error and classical design of experiments (DoE) to AI-DoE—using fewer experiments to reach better conditions faster.

Modern chemical experiments are costly and multidimensional. While many teams rely on experienced intuition and iterative testing (like one-factor-at-a-time (OFAT) approaches), these methods are best suited for narrow, low-dimension questions. They inevitably break down when faced with the complex interdependencies of a multi-variable system. In today’s high-dimensional landscapes, we have moved beyond the reach of intuition and can no longer rely on it to navigate modern chemical complexity.

In this guide, we unpack how chemists can evolve from old-school planning to an AI-augmented workflow, using tools like Atinary’s SDLabs to tie it all together.

In Traditional DoE, the goal is Characterization. You want a high-fidelity map of the entire neighborhood so you understand how every variable affects the outcome. In AI-DoE, the goal is Discovery. You don’t care about mapping the “valleys” (poor yields); you only want to find the “peak” (the optimum) as fast as possible.

Traditional methods like trial-and-error are slow and expensive. In chemistry R&D, each experiment can take hours or days of lab time, plus reagents cost. OFAT experiments compound this cost: for n factors, testing them one at a time can require far more runs than a multi-factor design. Moreover, chemical systems are inherently interactive—an adjustment in one variable (say catalyst loading) often changes how another (temperature) behaves. If these cross-effects aren’t captured, optimizing conditions becomes a game of luck.

Both classical DoE and AI-DoE address this problem, but they approach it very differently.

The systematic planning of DoE allows us to map the effects of controllable factors and their interactions with a minimal number of runs. For example, a 2-level full factorial design could vary temperature and pH at high and low settings, requiring just four runs to estimate each main effect and the interaction. Response Surface Methodology (RSM) designs extend this by including midpoints or center runs to model curvature.

A typical DoE workflow follows four steps:

Finally, chemists identify the predicted optimum conditions from the model—often confirmed by a few additional confirmation runs.

To illustrate: imagine optimizing a catalytic coupling. A scientist might start with a fractional factorial DoE to test four factors (e.g., ligand, base, temperature, solvent) at high/low levels. The analysis might reveal a critical interaction—perhaps a specific ligand only performs at high temperatures. A response surface DoE then narrows in, adding midpoints to model the “hill” of yield. While effective, this two-stage process is manual, time-consuming, and assumes the “hill” doesn’t shift unexpectedly.

As ASQ notes, DoE’s key advantage is that it “allows for multiple input factors to be manipulated, determining their effect on a desired output” and “identifies important interactions that may be missed when experimenting with one factor at a time”. In many process and formulation problems, DoE still makes sense. It gives an interpretable model and can be fully planned in advance when experiments are cheap and the factor space is small, like a dozen runs for 3 factors.

AI-DoE refers to closed-loop optimization, typically powered by Bayesian Optimization (BO). Instead of fixing a set of runs in advance, AI-DoE uses a machine-learning model to continuously learn from data and suggest the next experiments.

In a typical cycle, an initial set of experiments is run (chosen by some simple scheme, e.g. random or space-filling). The results train a surrogate model (often a Gaussian process) that predicts the objective (yield, conversion, etc.) across the untested space and quantifies uncertainty at each point. An acquisition function then balances:

The next experiment is chosen by maximizing this acquisition function, effectively seeking the condition expected to yield the most new information or improvement. This loop continues: after each batch of runs, the model updates and picks the next run. BO is designed to be sample-efficient. By explicitly modeling uncertainty and not repeating what it already “knows,” BO finds the optima in far fewer trials than exhaustive or grid searches.

DoE (left) vs. Bayesian Optimisation (right). DoE’s goal is understanding: it explores the parameter space systematically but non-adaptively, using pre-determined sample points that provide broad coverage. BO’s goal is optimisation: it learns and adapts its sampling after each iteration, using a model-guided search that focuses on the most promising regions.

In one study for a JAK2 inhibitor synthesis, researchers found that while traditional DoE provided a structured starting point, BO eventually outperformed it by adaptively targeting promising conditions in real-time. Instead of two separate, static stages, the AI-DoE Discovery Engine treated the entire process as one continuous, intelligent search.

In practice, many use a hybrid approach. A chemist might first use DoE to screen a wide space or eliminate obviously bad variables (for instance, testing which of 10 catalysts have any effect). Then, once the focus is on a smaller subspace, AI-DoE takes over to finely optimize conditions. This two-stage strategy leverages DoE’s speed for factor reduction, then BO’s precision for final tuning.

Theoretical efficiency is one thing; lab-proven performance is another.

In tackling a global economic and sustainability challenge, scientists from ETH Zurich (SwissCat+) utilized Atinary’s SDLabs AI Platform to optimize a catalyst formulation for CO2-to-methanol conversion. This high-dimensional problem involved 11 parameters and over 20 million possible combinations, a landscape far too vast for traditional trial-and-error.

By synthesizing and testing only 144 catalysts—occupying only 0.00072% of the design space—SDLabs delivered a 1,000× acceleration, replicating 100 years of traditional catalysis R&D in just six weeks.

The use of Bayesian Optimization through SDLabs navigated this complexity by simultaneously optimizing across four distinct objectives:

Read more about this study: publication link

Many daily essentials, from detergents and fragrances to pharmaceuticals, rely on the industrial chemical reaction, hydroformylation. However, this process often depends on Rhodium, which is among the scarcest and most expensive metals on Earth. For process chemists, maximizing efficiency while minimizing Rhodium loading requires navigating a complex 7D search space where temperature, pressure, and catalyst concentrations interact in non-linear ways.

In collaboration with Atinary, the chemists at dsm-firmenich utilized AI-DoE to find optimal reaction conditions by iteration 14 of a 22-iteration campaign. Running in batches of four, the closed-loop approach successfully reduced reaction time and total cost, proving that AI-driven insights can reconcile economic constraints with high-performance chemistry.

Read more about this study: publication link

AI-DoE is often framed as “fewer experiments to find the optimum.” That’s true, but it understates the full value. A well-run AI-DoE campaign delivers four distinct assets that compound over time.

The primary promise: AI-DoE reaches optimal or near-optimal conditions with a fraction of the experiments that classical DoE or OFAT would require. By surgically targeting the most informative regions of the search space rather than covering it uniformly, BO collapses weeks of screening into days. In the ETH Zurich catalyst study, only 144 experiments out of 20 million possible combinations were needed—a 1,000× acceleration.

BO doesn’t just hand you the best point and stop. After each campaign, the surrogate model provides a local map of the response surface around the optimum. You can read off how sensitive your yield is to small shifts in temperature, pressure, or catalyst loading. This local understanding answers practical questions that matter for scale-up: “How tight do I need to control pH?” or “What happens if my supplier changes the ligand purity by 2%?” Classical OFAT experiments rarely yield this kind of robustness insight.

Every AI-DoE campaign leaves behind a trained surrogate model—essentially a digital twin of your chemistry within the explored region. This model can be queried without running another experiment: predict performance at new conditions, estimate uncertainty, or simulate “what-if” scenarios. As campaigns accumulate, these models form a reusable knowledge base that accelerates future projects on related chemistry.

Unlike a DoE regression model that is valid only within the original factorial grid, a BO surrogate adapts its complexity to the data and can represent non-linear landscapes that a polynomial model would miss.

Perhaps the most underappreciated gain. Human experimentalists—no matter how experienced—carry implicit biases about which conditions “should” work. We default to familiar catalysts, safe temperature ranges, and established solvent systems. BO has no such preferences. Its acquisition function will propose experiments in regions a chemist might never have considered, as long as the model suggests potential. This systematic exploration of unconventional corners of the design space is where breakthrough discoveries live.

The ETH Zurich campaign, for instance, identified high-performing catalyst compositions that were non-obvious to domain experts. AI-DoE doesn’t replace the chemist’s intuition—it complements it by ensuring that intuition is tested against data, and that unexplored territory isn’t dismissed prematurely.

AI-DoE is, at its core, a tool built for the scientist. It doesn’t replace expertise—it amplifies it. By taking over the tedious, error-prone work of deciding what to try next, it frees chemists to focus on what they do best: asking the right questions, interpreting results with domain knowledge, and making the strategic calls that shape a research program.

More than efficiency, AI-DoE gives scientists the ability to go beyond their own bias. Every experimentalist—no matter how seasoned—has blind spots: familiar catalysts, comfortable temperature ranges, established protocols. AI-DoE systematically explores the regions we would have overlooked, and that is precisely where the most surprising discoveries emerge.

The results speak for themselves. Optimizing 11 parameters across 20 million combinations in six weeks. Finding optimal conditions by iteration 14 of a 22-run campaign. These outcomes were simply not achievable with traditional methods—not in the same timeframe, not with the same resources, and not without the closed-loop intelligence that adapts after every experiment. AI-DoE makes the previously impossible routine.

The scientist remains at the center: defining objectives, setting constraints, steering the search, and validating results. AI-DoE is the instrument that extends their reach—turning a handful of well-chosen experiments into insights that would have taken years to uncover by hand.

Visit our Use Cases section to learn how Atinary’s SDLabs AI Platform has accelerated R&D and process development across life-, chemicals-, and materials-science.