LLM converts literature into equations predicting defects in high-speed laser welding

Doing a lot with a little: LLM-based equations explain laser welding defects

September 22, 2025

A Penn State team has built an integration framework that turns sparse, mixed-format research into working equations for high-speed laser welding. Instead of waiting for massive new datasets, the approach blends minimal fresh experiments with information pulled from existing literature to predict process outcomes-especially defects-using an LLM-based workflow.

The result: quick, testable equations that help labs diagnose and reduce failure modes in precision welding for applications like EV fuel cells. The study is available online and scheduled for the October issue of the International Journal of Machine Tools and Manufacture.

The data bottleneck: not enough numbers, too much text

Traditional equation development is slow and data-hungry. Researchers either generate 1,000+ measurements themselves or manually extract and reconcile numbers buried across prior papers. That work often stalls because legacy results are described in text and figures rather than clean tables.

"With our model, we can simply input the literature data and substantially speed up the process," said doctoral candidate Zhengxiao Yu. The framework makes mixed, text-first literature practically usable without months of manual curation.

How the framework works

The team combined 48 datasets: five from new experiments and 43 from published studies. Their own experiments-covering metal types and defect observations like humping-seeded candidate equations that map variables to outcomes. They then built a rubric that directs the LLM to scan papers, identify relevant parameters, convert text-based descriptions into numeric form, and recommend equations most likely to hold under varied conditions.

Selection isn't one-click. The system generates about 10 candidate equations in a minute. Researchers score and rank them using a rubric, then pick the top performer. "We use the physical parameters of the weld like melt velocity, thermal conductivity and density to determine how the equation is going to be applied," said professor Jingjing Li. "This still requires a lot of domain knowledge from researchers, but could be built upon, and keeps our equations standardized across different materials and physical properties."

Targeting a persistent defect: "humping"

One focus is humping-a bead irregularity that often appears when travel speed is too high. According to doctoral candidate Zen-Hao Lai, the derived equations help explain when humping emerges by linking material properties and process settings to the flow and solidification behavior behind the defect.

This matters for production: if you can quantify how melt velocity, heat input, and material properties interact, you can set windows that maintain throughput while avoiding humping. The equations also support scenario testing without rerunning full experimental matrices.

Generalizing across metals and process windows

A key advantage is reusing prior experiments even when materials or machine speeds differ. By centering on physical parameters rather than fixed setups, the framework adapts insights from one context to another. That reduces the penalty of heterogeneous literature and lets teams make use of legacy work that previously didn't transfer well.

What this means for research teams

• Faster equation discovery: Go from hours per equation to minutes, then validate the best candidates.
• Better use of existing papers: Convert text- and figure-heavy reports into numeric inputs without manual retyping.
• Cross-study leverage: Apply insights across metals, speeds, and power settings using standardized physical variables.
• Defect-centric modeling: Quantify conditions that lead to humping and similar failure modes, and set practical process windows.
• Human-in-the-loop rigor: Keep expert oversight while moving the heavy lift of literature parsing and candidate generation to the LLM.

Scope beyond laser welding

The team plans to extend the framework to other manufacturing processes, including additive manufacturing. Anywhere you have limited new data but a deep literature base, this approach can accelerate model-building and reduce the gap between published research and shop-floor validation.

Citation

Derivation of physical equations for high-speed laser welding using large language models. International Journal of Machine Tools and Manufacture (2025). https://doi.org/10.1016/j.ijmachtools.2025.104320

Further resources

If your group is building LLM-assisted tooling for R&D and process engineering, you may find these training pathways useful: Complete AI Training - Courses by job.