What Phases is

The problem it solves#

You want to know what microstructure comes out of a particular cooling schedule, where the transformation curves sit for a candidate grade, or how large a weld nugget forms at a given current. Normally that means a solver license, a mesh, an input deck and half a day of setup before you learn anything.

Phases removes the setup. You describe the experiment in plain language. The agent resolves the material, configures the matching numerical solver, runs it, streams progress back, and attaches the plots, diagrams and stored files to the conversation.

These are real solvers#

This matters more than anything else on this page, so it is worth being blunt about it.

Phases covers four simulation domains: heat treatment and phase transformation, CCT diagram generation, resistance spot welding (RSW), and virtual Gleeble thermomechanical testing. Each one runs a real numerical solver on the server — these are not language-model approximations of physics.

The language model's job is narrow: understand what you asked for, pick the right tool, fill in its parameters, and report back what the tool actually returned. It is instructed never to claim a simulation ran, a diagram was produced, or a file was saved unless the corresponding tool was called and returned a result. Numbers in the reply come out of the solver, not out of the model.

That boundary is also why Phases refuses some questions. It will not tell you a weld preheat requirement, a toughness value, a quench-cracking risk, a solute retention or precipitation outcome, or an age-hardening response from composition and phase fractions alone. Those need calculations or test data the product does not have.

The four built-in simulators#

Two honest caveats belong right here rather than three pages later.

TTT (isothermal) diagrams are not produced by the CCT chat tool. Asking for one there returns an explicit "TTT transformation diagrams are not implemented yet" rather than a guess. Only continuous-cooling diagrams run from chat.

RSW is limited to the grades that have validated bulk electrical, thermal and mechanical properties: DC01, DC03, DC04, DC05, DC06, HC260LA, HC300LA, HC340LA, HC380LA, HC420LA, DP450, DP500, DP600, DP780, 20MnB5, 22MnB5, 20MnB8, 25MnB5, 29MnB7 and EN13261. Other grades are refused, not approximated. See what each simulator supports.

Each simulator has its own page: Heat treatment, CCT and TTT, Resistance spot welding and Virtual Gleeble.

The fifth tab is Custom#

The simulator strip carries five entries: HT Sim, CCT, RSW, V-Gleeble and Custom. Custom lists the simulators your own workspace has authored and published, and they run like any of the four built-ins.

A workspace simulator is built from chat rather than from a code editor: the agent proposes a design, drafts a versioned manifest out of approved operators, validates it, test-runs it, and an admin publishes an immutable version. Twelve of the chat tools exist for workspace simulators alone — nine for authoring and three for listing, inspecting and running what has been published. Publishing makes a version available; it is not a claim that the physics inside it is right, and Phases will not describe an experimental workspace simulator as validated. See Custom simulators and Running a custom simulator.

Custom is available where Simulator Studio is enabled, which is the default for every workspace today; an admin can switch it off, and the tab then disappears rather than erroring. The same is true of the four built-ins individually. See Workspace limits and features.

An older Spot Weld Tester (SWT) scaffold is no longer offered as a tab. It survives only as a label so historical run rows can still be named.

Materials#

The built-in catalog holds about 2,060 grades across 43 families and 68 alloy-system groups — structural steels, engineering steels, tool steels, stainless grades, cast irons and more. That is not a list worth printing in a manual, and the agent does not work from one: when you ask what is available it calls the live catalog and answers from what is actually there.

Grades you are likely to recognise — a handful out of thousands — include 42CrMo4, 16MnCr5, 100Cr6, C45E, 34CrNiMo6, 30CrNiMo8, 20MnCr5, 4140, 4340, 8620, 52100, H13, S355J2G3, DP600 and 22MnB5. Cast irons are largely composition-named rather than trade-named.

Nearly every grade carries alternative names, up to four each, so you can ask for a steel the way your drawing spells it. 316L also answers to 1.4404, X2CrNiMo17-12-2 and S31603; EN designation, DIN Werkstoffnummer, AISI/SAE number and UNS number resolve to the same card. Naming or describing the steel is the normal move — not asking for the whole list, which above 40 grades comes back as a summary by type and family with counts rather than thousands of names. See Finding your steel.

The transformation data behind those cards comes from published atlases and producer datasheets rather than from the model. The largest contributors are HTDG 2008 (380 grades), the NIMS database 2007 (194), Cias/Climax Molybdenum 1977 (120), IRSID 1974 (85) and the ASM Atkins atlas 1980 (70), alongside Böhler, Metal Ravne, Rose/Wever/Peter, Saarstahl, Deutsche Edelstahlwerke and others. Families and data sources sets out which family came from where.

Coverage is not uniform. Every grade in the catalog carries a verified card, but seven of them are not marked kinetics-ready, and RSW accepts only its own twenty. The capability matrix is the page that tells you, per simulator, whether your steel will actually run.

You can also supply your own chemistry. Phases identifies the closest named grade, tells you which one it is, and asks you to confirm before it creates anything — because a chemistry you typed is not a calibrated material. That gate is explained in Using your own chemistry.

Workspaces can save custom grades to a shared library. See Workspace materials and The material catalog.

Other things it computes#

Phases also predicts flow stress: per-phase and composite true-stress/true-strain curves for a material at a given ASTM grain size, returned with the model coefficients and their provenance. It runs behind its own confirmation card, the same way a simulation does. It is a prediction, not an applied material card — reviewing and applying a revision happens on the material's Flow stress tab.

Once you have runs, the agent can compare them and run bounded analysis recipes over their stored data instead of you exporting everything by hand. See Compare and analyse runs.

When no allowlisted recipe fits, it can fall back to bounded Python over up to four of your own runs, in a network-disabled sandbox. It can search the web when a datasheet or a published curve is the missing piece. And it can plan a bounded multi-trial simulation study whose trials run under one approved plan. Each of those is a separate workspace feature an admin can switch off.

What a session looks like#

You state a grade and process conditions in prose. The agent replies with an editable parameter card summarizing what it understood. You adjust anything you disagree with and press the run button. The simulation runs, streaming live progress. Results arrive as plots and diagrams attached to the conversation, stored under a numbered run you can return to later.

Simulations run on the server rather than inside the browser tab, so closing the tab does not stop one. While a simulation is running covers what you see, how it reconnects and how long a run has to finish.

There is no default material anywhere in that flow. "Run a simulation" or "surprise me" gets a question back, never a run. That is deliberate, and How the conversation works explains why.

Who it is for#

Metallurgists and process engineers who know steel and do not want to learn a simulation package to answer a bounded question. You need to know what you are asking; you do not need to know how the solver is wired.

You can try it without an account. Guest sessions are limited — 10 messages per 24 hours by default — and guests get only the chat itself, without the run browser, materials pages or history. Signing in adds persistent chat history, saved runs and artifact storage. See Guests and limits.

What Phases does not do#

• It does not replace review. Results are numerical simulations and, like any model, depend on input assumptions and material data quality. Do not present Phases output as validated engineering data without reviewing it. The chat itself carries the same warning under the composer: "Phases can make mistakes. Check important info."
• It does not label a cooling rate as a water, oil or air quench. A number is a number unless you supplied that mapping and the solver models it.
• It does not report per-phase hardness. Predicted hardness is a bulk, mixture value.
• It does not accept file uploads in chat, and it does not expose an unauthenticated public API for third-party simulation runs. The one supported programmatic surface is an authenticated MCP connection — see Use Phases from an MCP client.

The full list lives in Known limits. The evidence behind the physics is in How we validate, with the detail split into hardness validation and RSW validation.

Where to start#

Run something. Your first simulation takes about five minutes and uses a prompt you can paste as-is. Unfamiliar terms are defined in the Glossary.