Investment in materials and chemicals AI has attracted major focus, but the real industry breakthrough is the shift away from large language models (LLMs) to graph neural networks (GNNs), which are proving far more accurate for molecular prediction and materials design. Dave Rutledge, Head of Global Development at Tokyo deep-tech CrowdChem, explains why GNNs are outperforming LLMs in chemical R&D, how the industry is responding, and what it means for manufacturers who haven’t yet adopted AI.
For decades, chemical research and development (R&D) has largely relied on a time-tested but costly model: trial and error.
Scientists and engineers iterate through experiments, testing different material formulations, coatings, or composites, often guided by intuition, human expertise, and incremental tweaks. This process, while foundational for many breakthroughs, is slow, wasteful, and expensive.
Today, AI is fundamentally transforming that paradigm. Rather than relying on blind experimentation, companies can now use predict-and-verify workflows: AI models suggest promising candidates, guide which experiments to run, and help validate them which dramatically reduces the number of failed trials. This shift is not just theoretical but is already unlocking major gains in areas like energy storage, composites, and surface treatments.
Traditional R&D typically depends on human-led experimentation. Researchers formulate a material, run tests, analyse results, adjust, and repeat. Each cycle takes time, resources, and often large volumes of materials, especially in sectors like coatings or advanced composites.
This approach has three big drawbacks:
High cost: Physical experiments consume chemicals, energy, lab time, and manpower.
Long timelines: Iterative cycles mean it may take months or years to converge on optimal formulations.
Wasted resources: Many experiments fail or only yield incremental improvements.
In many sectors, this traditional method has barely changed in half a century. AI changes this fundamentally. Rather than testing everything in the lab, AI-driven models can accurately predict which material formulations are likely to work, filter out unpromising ones, and guide experiments more intelligently.
The predict-and-verify workflow uses AI to streamline R&D by guiding experimentation rather than relying on guesswork. First, models are trained on existing data, such as past lab results and material properties, to learn how different parameters influence performance.
They then predict which formulations or process conditions are most likely to meet specific targets, from durability to conductivity. Researchers run a small, focused set of experiments to validate these predictions, and the results feed back into the model, sharpening its accuracy over time.
This continuous loop significantly reduces the number of experiments required while accelerating discovery.
For example, in battery R&D, discovering new materials for electrodes or electrolytes traditionally meant synthesising and testing dozens (if not hundreds) of variants. AI models can predict which combinations of chemical components (e.g., salts, solvents, additives) are likely to deliver performance targets such as higher energy density or longer cycle life, reducing the number of expensive physical tests.
It’s tempting to imagine dropping a powerful LLM into lab R&D and having it “figure out” new materials. However, in reality, general-purpose language models are not well-suited to physical science.
• LLMs treat numbers as text, making them unreliable for high-precision regression required in chemistry.
• Longer experimental histories break LLM accuracy, as context length increases.
• Fine-tuning large LLMs is costly and inefficient, limiting rapid iteration.
• LLMs flatten scientific workflows into text, losing the hierarchy between materials, steps, conditions, and results.
• LLMs are better for extraction than prediction, helping convert documents into structured data but not generating accurate scientific outputs.
Instead, Graph neural networks (GNNs) represent one of the most important, and least publicised, breakthroughs in scientific AI. Unlike LLMs, GNNs view data as networks of interconnected components. GNNs are more “chemistry-native”.
GNNs preserve the true structure of chemical processes, using directed trees that reflect materials, conditions, and steps. They combine multiple data types including numeric values, text descriptions, and molecular information into one coherent representation. Because GNNs align with the physical structure of matter, they can generate meaningful predictions even when data is scarce, making it a critical advantage in science-heavy industries.
The transition from trial-and-error to predict-and-verify is more than a technical upgrade. It represents a cultural shift in R&D. AI will not only accelerate innovation, but also democratise it. Smaller companies with fewer resources can compete by leveraging predictive models to guide their experiments.
The future of manufacturing R&D will be defined by intelligent experimentation, where machines and humans collaborate in a tight loop of prediction, verification, and refinement. LLMs will continue to reshape communication and knowledge work. But they are not the models that will solve our hardest engineering and climate challenges.
Crucially, AI is not here to replace scientists or engineers. By handling repetitive data processing and narrowing the field of promising candidates, AI allows scientists to spend more time doing science, and engineers to focus on engineering. Rather than automating people out of the process, AI amplifies human expertise and removes bottlenecks that prevent teams from working at their full creative and technical potential.
Manufacturing R&D has long been stuck in a cycle of slow, resource-intensive trial and error. With AI, that’s changing. By shifting to a predict-and-verify model, companies can radically reduce waste, cost, and time-to-market and accelerate innovation in critical sectors.
The most powerful applications arise when domain experts and data scientists work together, using specialised models tailored to the physical, chemical, and structural properties of materials. The promise of AI in this context isn’t just about automation, it’s about smarter experimentation, more efficient discovery, and more sustainable manufacturing.
We are entering a new era where R&D is not measured in failed trials, but in validated predictions. The companies that embrace this approach will lead the next wave of industrial innovation.
CrowdChem has published a major academic paper on the GNN model here.
The research is based on one of the largest graphical chemical process datasets in the world, with over 700,000 structured process graphs extracted from more than 9,000 scientific documents. This scale allowed CrowdChem’s team to directly compare LLM-based approaches with the new graph neural network (GNN) architecture designed specifically for materials and chemical experimentation.
