⚡ electrical engineering

AI-assisted finite element modelling of the electron beam and TIG welding processes of low-cost titanium alloys

This study demonstrates that an AI-assisted workflow, utilizing large language model agents to replace conventional CALPHAD computations, successfully predicts the distinct phase compositions of electron beam and TIG welded low-cost Ti-2.8Al-5.1Mo-4.9Fe titanium alloys by coupling finite element thermal simulations with AI-generated continuous-cooling-transformation diagrams.

Original authors: Roman Selin, Serhii Akhonin, Valeriy Bilous

Published 2026-08-13

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Roman Selin, Serhii Akhonin, Valeriy Bilous

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a chef trying to bake the perfect cake, but instead of flour and sugar, your ingredients are metals like titanium. Titanium is a super-ingredient: it's incredibly strong but surprisingly light, making it a favorite for building airplanes, rockets, and even medical implants. However, the traditional recipe for the strongest titanium cakes is expensive because it uses a rare spice called vanadium. Scientists have been trying to create a "budget-friendly" version of this recipe by swapping vanadium for cheaper ingredients like iron and molybdenum. But there's a catch: when you melt and re-freeze metal to join two pieces together (a process called welding), the heat changes the metal's internal structure, just like how baking changes the texture of a cake. If the metal cools down too fast or too slow, the "cake" might turn out brittle or weak. To predict this, scientists usually need super-complex math software that acts like a crystal ball, telling them exactly how the metal's tiny internal grains will rearrange as it cools. This paper asks a bold question: Can we replace that expensive, complicated crystal ball with a smart computer chatbot to get the same answer?

The researchers in this study decided to test this idea on a specific low-cost titanium alloy (a mix of titanium, aluminum, molybdenum, and iron). They wanted to see how two very different ways of welding this metal would change its final structure. One method, called Electron Beam Welding (EBW), is like using a laser pointer: it's incredibly focused, hot, and fast, punching a deep, narrow hole through the metal. The other method, TIG welding, is more like a campfire: it's a wider, gentler heat source that creates a shallow, broad pool of melted metal. The team built a computer simulation to act as a "digital twin" of the welding process, tracking exactly how fast the metal cooled down in different spots. Then, instead of using the traditional, heavy-duty math software, they used two AI agents (smart computer programs) to predict the result. The first agent, the "CCT-Agent," acted like a meteorologist, drawing a map of how the metal would change as it cooled. The second agent, the "Phase-Agent," acted like a translator, reading the cooling speed from the simulation and telling them exactly what the metal's internal structure would look like.

The results were fascinating. The simulations showed that the two welding methods created completely different "cakes." The fast, focused Electron Beam weld cooled the metal down at a blistering speed of about 50 degrees Celsius per second. Because it cooled so quickly, the metal didn't have time to rearrange its internal grains into a new shape; it stayed in a "frozen," high-energy state called metastable beta. In contrast, the slower, wider TIG weld cooled down much more gently, at about 10 to 15 degrees Celsius per second. This slower pace gave the metal plenty of time to rearrange itself, forming a significant amount of a different, stiffer structure called diffusional alpha. The study found that the AI agents were able to draw these cooling maps and predict the final structures just as accurately as the traditional, expensive software, but much faster and without needing special licenses. Essentially, the researchers proved that for this specific budget-friendly titanium, the choice of welding tool isn't just about how the pieces fit together; it's a choice between two completely different internal personalities for the metal, and a smart AI can now help engineers predict which one they'll get before they even pick up a welding torch.

Technical Summary: AI-Assisted Finite Element Modelling of EBW and TIG Processes in Low-Cost Titanium Alloys

Problem Statement
Low-cost $\beta$ -titanium alloys, designed to replace expensive vanadium with iron and molybdenum (e.g., Ti-2.8Al-5.1Mo-4.9Fe), offer significant potential for weight-critical aerospace and marine structures. However, their weldability is poorly understood compared to the dominant Ti-6Al-4V alloy. A critical challenge in welding these metastable- $\beta$ alloys is predicting the weld-zone phase composition, which is highly sensitive to thermal history. Traditional prediction workflows rely on Finite Element (FE) thermal modelling coupled with CALPHAD (CALculation of PHAse Diagrams) to generate Continuous Cooling Transformation (CCT) diagrams. This conventional approach is often a bottleneck, requiring specialized thermodynamic databases, licensed software, and significant expertise, making it slow and difficult to apply when screening multiple low-cost compositions.

Methodology
The study employs a coupled computational workflow integrating FE thermal simulation with an Artificial Intelligence (AI) agent-based microstructure prediction system.

Thermal Modelling: Transient 3D FE analysis was performed on 6 mm plates of Ti-2.8Al-5.1Mo-4.9Fe to simulate two distinct welding processes: Electron Beam Welding (EBW) and Tungsten Inert Gas (TIG) welding. The models utilized temperature-dependent thermophysical properties and latent heat formulations. EBW was modeled as a concentrated conical/cylindrical-Gaussian source (keyhole mode), while TIG was modeled as a diffuse double-ellipsoidal source (conduction mode). The simulations extracted zone-resolved cooling rates, specifically over the 800–500 °C interval relevant to $\beta \to \alpha$ transformation.
AI-Driven CCT Construction (CCT-Agent): To replace the CALPHAD step, a Large Language Model (LLM) agent, the "CCT-Agent," was developed. Operating on a locally hosted, open-weight model (gpt-oss, 120B parameters) via the Ollama API, the agent constructs CCT diagrams directly from alloy chemistry. It utilizes metallurgical anchors (e.g., $\beta$ -transus, molybdenum equivalent) and applies Johnson-Mehl-Avrami-Kolmogorov (JMAK) kinetics with Scheil additivity to define transformation boundaries. The workflow is deterministic (temperature 0, fixed seed) and schema-validated to ensure reproducibility.
Phase Prediction (Phase-Agent): A second LLM agent, the "Phase-Agent" (running on a 20B parameter variant), integrates the FE-derived cooling rates against the CCT-Agent's generated diagrams. It calculates the resulting phase fractions (diffusional $\alpha$ vs. retained $\beta$ ) for specific locations within the weld and heat-affected zone (HAZ).
Validation: Before application to the low-cost alloy, the CCT-Agent was validated against CALPHAD-derived diagrams and literature data for three alloys: Ti-6Al-4V ( $\alpha+\beta$ ), VT19 (near- $\beta$ ), and the primary Ti-2.8Al-5.1Mo-4.9Fe (metastable- $\beta$ ).

Key Results

CCT-Agent Validation: The AI agent successfully reproduced CALPHAD-derived transformation boundaries within engineering tolerances across the stability spectrum. For Ti-6Al-4V, VT19, and Ti-2.8Al-5.1Mo-4.9Fe, the agent matched $\beta$ -transus temperatures within approximately 30 °C and correctly identified the ordering and magnitude of critical cooling rates required to retain the $\beta$ phase.
Thermal Divergence: The FE models confirmed distinct thermal signatures. EBW produced a deep, narrow fusion zone (depth-to-width ratio ≈ 2.5) with high cooling rates of approximately 50 °C s⁻¹ in the fusion zone. Conversely, TIG produced a wide, shallow fusion zone (depth-to-width ratio ≈ 0.38) with significantly slower cooling rates of 10–15 °C s⁻¹ in the fusion zone and ~6 °C s⁻¹ in the HAZ.
Microstructural Prediction: The coupled workflow predicted starkly different phase compositions for the same alloy based on the welding process:
- EBW: The rapid cooling (~50 °C s⁻¹) bypassed the diffusional $\alpha$ -start nose of the CCT diagram. The fusion zone and narrow HAZ remained predominantly metastable $\beta$ .
- TIG: The slower cooling (~10–15 °C s⁻¹) allowed the thermal trajectory to intersect the transformation boundaries. The weld developed a graded microstructure with a retained $\beta$ core surrounded by significant diffusional $\alpha$ formation in the weld metal and HAZ.

Significance and Claims
The paper demonstrates that LLM agents can fully replace the CALPHAD computation step within a welding simulation loop. By generating CCT diagrams and phase maps directly from composition and cooling rates, the two-agent workflow (CCT-Agent and Phase-Agent) offers a fast, accessible, and reproducible alternative to traditional thermodynamic modelling. This capability is particularly significant for screening low-cost alloy compositions where establishing reliable CALPHAD databases is resource-intensive.

The study concludes that for the low-cost alloy Ti-2.8Al-5.1Mo-4.9Fe, the choice between EBW and TIG is effectively a choice of as-welded phase balance. EBW yields a uniform metastable- $\beta$ structure suitable for subsequent ageing, while TIG yields a structure with pre-formed diffusional $\alpha$ , necessitating different post-weld heat treatment strategies. The authors note that while the CCT-Agent is validated against established methods, the specific phase predictions for the low-cost alloy welds have not yet been experimentally verified via metallography or X-ray diffraction, marking this as the essential next step. The workflow is presented as a tool for operationalizing established metallurgical relationships rather than inventing new thermodynamics.

More like this