Chemistry LLMs in the Real World: What a Discovery Call Taught Me About AI in Chemical R&D

Everyone’s talking about Chemistry LLMs — models that generate novel molecules, predict material properties, and accelerate drug discovery. But when I sat down with data scientists from a global specialty chemicals company, the conversation went in a very different direction. This post captures what enterprise chemistry teams actually need versus what the AI hype cycle promises.

The Problem

The narrative around Chemistry LLMs goes something like this: train a large language model on SMILES strings (a text-based molecular notation), fine-tune it on your proprietary formulation data, and let it generate novel molecules that meet your target properties. Papers from research labs show impressive results. Startups raise funding on this vision.

But enterprise chemistry teams live in a different world. They have decades of formulation data scattered across Excel files, local machines, and SQL databases. Their data scientists spend most of their time cleaning and structuring data, not training models. Their researchers already know what molecules they want to test — they need validation and ranking, not pure generation.

The gap between “Chemistry LLM generates novel molecules” and “enterprise chemistry team ships value” is enormous. And most of that gap has nothing to do with model architecture.

The Solution

The real opportunity in enterprise chemistry AI isn’t a better model — it’s a better data platform. Getting data from researchers, cleaning it, structuring it, sharing it across regions, and deploying predictions back into the workflow. That’s the bottleneck. A Chemistry LLM may come later, but it’s not the priority today.

This reframing matters for anyone building or selling AI solutions to manufacturing and chemicals companies. If you lead with “here’s a generative model,” you’ll get polite interest. If you lead with “here’s how to solve your data problem at scale,” you’ll get budget.

How It Works

The Discovery: More Advanced Than Expected

During the call, the data scientists revealed they had already tested multiple model families:

Model Family	Type	What It Does
ChemBERTa, SmilesBERT	Encoder (BERT-like)	Property prediction from molecular representations
Transformer generative	Decoder (GPT-like)	Molecule generation from SMILES
Diffusion-based	Generative	Alternative generation approach
Bayesian optimization	Classical ML	Formulation optimization (their preferred method)
XGBoost, neural nets	Classical ML	Predictive modeling for existing formulations

They weren’t starting from zero. They had a targeted, use-case-by-use-case approach — pick the most performant model for each specific class of molecules. No single global strategy.

The key finding: smaller, targeted models often outperform larger pre-trained foundation models for their niche chemical space. This is the opposite of what the LLM scaling narrative suggests, but it makes sense when your domain is highly constrained (small molecules with a definite number of atoms).

Data Is the Real Bottleneck

When I asked about their biggest challenge, the answer wasn’t “we need a better model.” It was data. Every aspect of data was painful:

Representation: How do you describe a molecule computationally? The quality of your molecular representation directly drives model performance. Their existing predictive system uses chemical names in a SQL database — no SMILES strings, no SDF files. Converting between representations requires domain expertise.

Collection: Data lives in Excel spreadsheets, on researchers’ local machines, in disparate databases. Getting data from researchers involves asking, emailing back and forth, waiting. There’s no centralized data lake.

Cleaning: Python scripts for data cleaning, rarely using cloud-based tools for data preparation. Data cleaning consumes the majority of their time. Each use case requires different cleaning and structuring approaches depending on the data volume and quality.

Cross-region sharing: Research centers in different countries working with different data subsets. Sharing data across regions is logistically painful.

This pattern repeats across every manufacturing company I work with. The AI conversation always starts with models and ends with data infrastructure.

Substitution Beats Generation

Here’s the insight that reframed the entire conversation: leadership doesn’t want to discover new molecules. They want to reuse existing ones in new applications.

The priority use cases break down like this:

• Substitution (highest priority): Find which existing chemicals in the portfolio can replace a target molecule. Think regulatory compliance — replacing restricted substances with approved alternatives from existing formulations and patents.
• Property prediction (high priority): Given a formulation, predict its properties. Reduce the number of physical experiments needed. Directly tied to time-to-market.
• Knowledge management: Capture tacit knowledge from senior researchers before they retire.
• Novel generation (lowest priority): Generate entirely new molecules from scratch. Few projects actually need this.

Researchers already have candidates in mind. They need a system that validates, ranks, and optimizes — not one that dreams up molecules from scratch. Generating SMILES strings sounds impressive in a paper, but it’s logistically difficult at enterprise scale when you need to actually synthesize, test, and certify what the model suggests.

Why Generative LLMs Face Skepticism

The data scientists were frank: they’re skeptical about the business value of generative LLMs for formulation work. Their reasoning:

1. Niche chemical space limits foundation models. Pre-trained Chemistry LLMs are trained on broad chemical datasets (PubChem, ZINC). The company’s specific polymer/adhesive/coating chemistry is a narrow slice. Transfer learning helps, but fine-tuning data is limited.

2. Bayesian optimization works better for formulation. When you’re optimizing a formulation with 10-20 ingredients within known constraint ranges, Bayesian optimization is more sample-efficient and interpretable than a generative LLM.

3. Small molecules are harder. Unlike drug discovery (large, flexible molecules), specialty chemicals often deal with small molecules where the number of atoms is tightly constrained. The generation space is much smaller, and brute-force enumeration can compete with learned models.

4. Logistics kills the ROI. Even if a model generates a promising molecule, someone has to synthesize it, test it, validate it against regulations, and qualify it with customers. The experimental validation cost dwarfs the computational cost.

The Sensitive Topic: Vendor Lock-in During RFP

An interesting signal emerged during the call. When the discussion touched on their existing predictive AI platform (in production, handling formulation predictions), the team became noticeably uncomfortable and declined to discuss details.

The reason became clear: they had an active RFP in progress with multiple system integrators bidding to build the next version. Any detailed technical discussion with a cloud provider during an open procurement could compromise the process.

Lesson for Solutions Architects: always check whether there’s an active RFP before diving deep into a customer’s existing production systems. If the customer can’t discuss something, don’t press — note it and work around it.

What Enterprise Chemistry Teams Actually Need from Cloud

Based on this and similar conversations across the manufacturing sector, here’s what actually moves the needle:

What They Ask For	What They Actually Need
”A Chemistry LLM”	A data platform that makes their existing models work better
”Generative AI for molecules”	Better data collection, cleaning, and representation pipelines
”Foundation model fine-tuning”	MLOps infrastructure to deploy and iterate on smaller, targeted models
”Novel molecule generation”	Substitution search across their existing portfolio
”One big model”	Per-use-case model selection and orchestration

The real AWS value-add is data platform + MLOps at scale: getting data from researchers into a structured format, building reproducible training pipelines, deploying predictions back into the R&D workflow, and enabling cross-region collaboration. SageMaker handles the ML lifecycle. A data lake on S3 with Lake Formation handles governance. Step Functions or MWAA orchestrates the pipelines.

A Chemistry LLM might sit on top of all this eventually. But without the data foundation, it’s a solution looking for a problem.

What I Learned

• Enterprise AI maturity is higher than you think — but not where you expect. The data scientists had tested ChemBERTa, SmilesBERT, diffusion models, and Bayesian optimization. They weren’t waiting for someone to explain transformers. Their bottleneck was infrastructure and data quality, not model awareness. Always do a proper discovery before assuming the customer needs education.

• “Data is the hardest part” isn’t a cliché — it’s the entire problem. In chemistry AI, molecular representation quality directly drives model performance. If your data lives in Excel files with chemical names instead of SMILES strings, no amount of model architecture innovation will help. The conversation should start with “show me your data pipeline,” not “which LLM are you using.”

• Substitution is more valuable than generation in industrial chemistry. The hype cycle focuses on generating novel molecules. But the highest-value use case for most chemicals companies is finding which existing chemicals in their portfolio can replace a restricted substance. This reframes the technical problem from generative modeling to similarity search and property prediction — a much more tractable problem.

• Active RFPs create invisible walls. If a customer suddenly can’t discuss their production system, an active procurement is likely the reason. Don’t push — it damages trust. Note the sensitivity, work around it, and revisit after the RFP closes.

Do It Yourself

• Key takeaways: Lead with data platform, not model architecture. Enterprise chemistry teams need data infrastructure (collection, cleaning, representation, cross-region sharing) before they need generative AI. Substitution and property prediction deliver faster ROI than novel molecule generation. Always check for active RFPs before deep-diving into existing production systems.

• Try it now:

  1. Run a proper discovery before proposing solutions — Use structured questions: What models have you tested? What data formats and volumes? What does success look like? You’ll often find the customer is more advanced than expected but blocked on infrastructure.
  2. Explore molecular ML on SageMaker — Start with Amazon SageMaker with Hugging Face to fine-tune ChemBERTa or MoLFormer on a public dataset like MoleculeNet. This gives you hands-on experience with the representation challenges before you pitch to a customer.
  3. Study the chemistry ML landscape — Key papers: ChemBERTa (molecular property prediction), MoLFormer (large-scale molecular representation), and the Molecular Sets benchmark. Understanding what these models can and can’t do will make your customer conversations dramatically better.