2026-07-28
Extracting ALD and ALE recipes from literature with LLMs
How can we use large language models (LLMs) in process development? The easiest answer is to just ask a question and get a response. Recent research from Argonne National Lab benchmarks exactly how far we can push models to give us accurate answers out of the box. Another study explores how well LLMs can be integrated directly with lab instruments.
But when applying these tools to actual process development, we hit a wall: how confident can we be in their answers?
LLMs are notoriously bad at citing real references, especially since many scientific papers are trapped behind paywalls. If a model hallucinates a process parameter, how do we double check it without spending hours digging through the literature?
The data bottleneck
If we want to push models to be genuinely useful for process development, whether for direct queries, fine tuning, or training machine learning frameworks like generative models for ALD workflows, we need a dataset built directly from peer reviewed literature.
The problem? There is simply too much to read. If you search for ALD keywords on OpenAlex, you will find around 30,000 articles. Reading and extracting parameters from them manually is impossible.
This is where we can flip the script. Instead of asking an LLM a process question, we can use LLMs to read those 30,000 papers for us, filter out the noise, extract the exact parameters we need, and save them in a machine readable format. A recent article highlighted a similar approach for extracting parameters for a specific process from a curated database.
Introducing ALPminer
To solve this, I built ALPminer, an end to end pipeline that automates this entire extraction process.
Here is how it works: First, by using OpenAlex the software creates a database of articles based on your keywords and downloads whatever is openly available. For papers restricted by paywalls, it makes the manual process as painless as possible. Just drop the downloaded PDF into a folder, and the pipeline automatically curates and links it.
Next, a lightweight LLM acts as a screener, reading through the stack to filter out papers that do not actually report process parameters. The relevant papers are then passed to a more powerful LLM. For each parameter we care about, the model is given a specific, constrained query to find and normalize that exact data point.
Finally, the pipeline compiles everything into a clean JSON database, ready for your next machine learning project or process search.
Beyond individual research projects, this framework makes it possible to build a large comprehensive database that can be shared across the entire scientific community. By continuously maintaining this database, new articles can be seamlessly added the moment they are published, creating an ever growing, up to date repository of process recipes for everyone to use.
Built for flexibility
I wanted this tool to be accessible, so ALPminer features a fully graphical web based interface. Every step of the pipeline is adjustable. You can swap out the choice of LLM, tweak the prompts, or change the target parameters.
It ships with two ready to use profiles: one for atomic layer deposition (ALD) and one for atomic layer etching (ALE). However, the underlying engine works for any process that relies on parameters. You can easily create and save custom profiles tailored strictly to your own needs.
ALPminer is open source and ready to use. You can download it from my GitHub repository. Just run the bash or command file for your operating system, and the web interface will open automatically. Everything runs locally, meaning your data and downloaded PDFs stay entirely on your own device.
Check out the README file for more details, and let me know how it works for your workflows!