AI models for chemistry often fail on compounds no one has tested because training data is scarce. In a Data Science Connect webinar, SandboxAQ's Ang Xiao explained how large quantitative models add physics-based synthetic data, with case studies in battery life prediction, paper mining, materials discovery and formulation.
Key takeaways
- Chemistry lacks training data. AI predictions outside explored chemical space are often wrong, Xiao said.
- LQMs add physics. First-principles simulations such as DFT generate synthetic data that models are trained or fine-tuned on.
- LLMs can mine the literature. SandboxAQ's Extract tool turned research papers into a materials data set with 90% reported precision and recall.
- Screening funnels save years. Combining surrogate and quantum chemistry models cut electrolyte discovery from years to days, according to Xiao.
- AI experiment recommenders cut lab work. Bayesian optimization reduced experiments by 60% in one formulation project.
Large language models learn from the internet's text. Chemistry and materials science don't have that kind of training data, which is why AI models in those fields often fail on compounds no one has tested. In a Data Science Connect webinar, Ang Xiao, PhD, Technical Lead of AI and Quantum Applications at SandboxAQ, explained how large quantitative models (LQMs) try to close that gap, and walked through case studies in batteries, data extraction, materials discovery and formulation. Brian Mink, co-founder and president of Data Science Connect, hosted.
Xiao said he worked in the chemical and materials science industry for more than two decades before leading AI and quantum applications for chemistry and materials science at SandboxAQ.
Why is materials R&D so slow?
Take a new battery material for electric vehicles that needs high energy density and must also be safe. "We have way too many choices," Xiao said, describing a space of possible material combinations far too large to test. In his experience, finding a viable candidate typically takes five to 10 years of synthesizing and testing thousands of materials, because the final product has to meet requirements for properties, performance, toxicity, environmental fate and cost.
Most labs work through a design, make and test cycle that leans heavily on human intuition to decide which leads are promising. "We should be able to do better," Xiao said. He described a platform that generates millions of candidate compounds, predicts their properties with AI or physics-based models on high-performance and cloud computing, and sends only the best to the lab.
What is a large quantitative model?
The core problem, Xiao said, is that AI has explored only a tiny fraction of chemical space. When a model predicts outside the region covered by its training data, "it's often wrong."
LQMs address this by adding physics. Simulations based on first-principles calculations, such as density functional theory (DFT), generate synthetic data points in unexplored regions. Models built on existing architectures, including transformers and graph neural networks, are then trained or fine-tuned on that data so they can predict energies and properties without rerunning expensive quantum chemistry simulations each time.
Asked whether a large technology company could do the same, Xiao said it could, but that domain knowledge matters. "Without domain knowledge, sometimes we find out the model training is not accurate," he said.
How can AI predict battery cycle life?
Standard cycle-life models rely on capacity and voltage curves from ordinary battery cyclers, which Xiao said lack accuracy. SandboxAQ instead starts with data from a partner's ultra-high-precision cycling equipment, then applies its own feature engineering and data transformations, including smoothing to reveal underlying trends.
Using differential capacity curves from only cycles five and six, the model achieved a mean absolute error of 46 cycles with an R-squared of 0.99, Xiao said, calling it a strong result for early prediction.
How do LLMs help extract chemical data from research papers?
SandboxAQ built a tool called Extract that uses LLMs such as GPT-4 to pull material properties out of thousands of research papers and turn them into training data. The workflow sources papers through publisher partnerships, cleans messy XML into readable text, splits and filters passages by units and criteria, and then runs a zero-shot extraction step. Each extracted value links back to the passage it came from.
In a project with the U.S. Army, Xiao said, the team started with more than 6,000 materials and narrowed them to 48 with complete property values. The model showed 90% precision and recall, with about 10 minutes of human setup time and three hours of GPU computing. A live demo showed the tool correctly pulling a yield strength value and its source passage.
What can LQMs do for materials discovery and formulation?
Xiao described several case studies, all reported by SandboxAQ:
- Battery electrolytes: A funnel combining fast surrogate models with more accurate quantum chemistry models narrowed millions of compounds to a few, reducing discovery "from years to just several days."
- Oxygen ion conductors: DFT-based screening of tens of thousands of materials identified candidates for energy storage devices such as fuel cells.
- PFAS: SandboxAQ scaled an advanced quantum chemistry method, incremental full configuration interaction, across 1 million CPU cores on AWS to calculate carbon-fluorine bond dissociation energies in PFOA, a first step toward breaking down so-called forever chemicals.
- Electrolyte formulation: An AI experiment recommender found an additive loading for silicon anodes that reduced gassing and maximized cycle life. In another project, Bayesian optimization cut the number of experiments by 60% compared with random search.
- Fast charging: For graphite and lithium iron phosphate 18650 cells, the recommender found the best charging protocol in 30 tests, cutting experiment time by 37%, Xiao said.
He closed by arguing that AI's biggest potential lies beyond language, graphics and video, in solving high-dimensional problems in chemistry, physics and biology.
More applied AI sessions ahead
Data Science Connect has 24 free live webinars on its 2027 schedule across the AI × Enterprise Panel Series, AI Demo Days, Agentic AI Showcases and the Plug & Play Series. See the webinar schedule. Leaders applying AI in research and operations can also meet in person at The AI Enterprise Conference on Tuesday, August 31, 2027, at Pier Sixty in New York City.
Frequently asked questions
What is a large quantitative model (LQM)?
As Xiao described it, an AI model for chemistry, physics or biology that is trained or fine-tuned on synthetic data from physics-based simulations, letting it predict beyond the limits of experimental data.
How are LQMs different from LLMs?
Xiao said SandboxAQ uses LLMs only to extract data from papers; LQMs rely on quantum chemistry simulations to generate data, then use architectures such as transformers and graph neural networks.
Can AI predict battery cycle life early?
Xiao said SandboxAQ's model, using high-precision data from cycles five and six, predicted cycle life with a mean absolute error of 46 cycles.
How does AI reduce the number of lab experiments?
Experiment recommenders using Bayesian optimization pick the most informative next test; Xiao reported a 60% reduction in experiments for an electrolyte project.
Watch: Reimagining R&D: The Game Changing Potential of Large Quantitative Models on YouTube
Event: Webinars · More coverage: Webinars insights
