Introduction

LatticeMind is an AI agent for computational materials science. You describe the calculation you want in plain language — “compute the band structure and density of states of silicon” — and LatticeMind carries out the entire workflow for you: it plans the sequence of calculations, builds the crystal structure, chooses sensible numerical parameters, writes the solver input files, checks them for physical and syntactic correctness, runs the calculation (on your laptop or on a remote HPC cluster), analyses the raw output, and produces a readable report with figures.

At its core LatticeMind drives the RESCU density functional theory (DFT) solver, and it speaks RESCU fluently: self-consistent field, band structure, density of states, structural relaxation, DFT+U, density functional perturbation theory (DFPT) phonons, and more. It also has an experimental backend for the NanoDCAL quantum-transport solver and a backend for VASP.

What makes LatticeMind different from a template generator is that it reasons about the science. It maintains an explicit, validated plan of the work; it wires multi-step workflows together correctly (feeding the converged density of an SCF run into a subsequent band-structure run, for example); and it checks its own work at every stage, repairing mistakes before a single core-hour is spent.

How LatticeMind works

A LatticeMind project moves through a fixed sequence of stages: a planner turns your request into an explicit plan, a structure builder constructs the geometry, a parameterizer chooses the numerical settings, an input builder renders the solver decks, a validator gates execution, an execution backend runs the job (locally or on HPC), and a reporter writes up the results. Each stage inspects the output of the previous one and either proceeds or asks for a repair, so errors are caught early rather than surfacing as a failed calculation hours later.

Fig. 1 The LatticeMind pipeline. A natural-language prompt is turned into a validated, executable project and finally a report. Validation can send the plan back for repair (self-healing) before anything is executed.

You stay in control throughout: solver execution is off until you enter /execute on. Use /cost to inspect the estimate and /launch-preview to inspect the exact local or remote launch before a job starts. For a stage-by-stage walkthrough — including the self-healing loop and the human-in-the-loop gates — see How LatticeMind works.

Supported solvers

LatticeMind binds one project to one solver. The three backends and their current maturity are:

Solver	Status	What it covers
RESCU	Primary, stable	SCF, band structure, density of states, projected/local DOS, structural relaxation, DFT+U, DFPT phonons and dielectric response, spin-polarized and non-collinear magnetism, alloys and defects.
NanoDCAL	Experimental	Quantum transport for two-probe devices (transmission, conductance, I–V, scattering states), effective mass, complex band structure, phonons, and ground-state properties. See NanoDCAL support (experimental).
VASP	Preview (beta)	Multi-stage plane-wave workflows (SCF, relaxation, band structure, DOS).

Note

NanoDCAL support is under active development. Its supported calculation types have each been run end-to-end during internal testing, but the surface is smaller and evolving faster than the RESCU backend. This manual documents RESCU first and treats NanoDCAL as an advanced, experimental topic.

Key features

• Plain-language interface. No input-file syntax to memorize. Prompts can be as terse as “SCF for fcc nickel” or as detailed as you like.

• Self-healing workflows. Every plan is validated before execution and automatically re-planned when a check fails. Failed attempts are archived, not discarded, so you can see what changed.

• Correct multi-step chaining. Densities and wavefunctions are handed from one calculation to the next automatically, with unique result paths per step.

• Three interfaces. An interactive terminal Dashboard implemented with the Textual terminal-interface framework and Rich terminal-rendering Python library, a browser-based Web UI, and a plain CLI for servers and notebooks — all driving the same engine. See The three interfaces.

• Bring your own AI model. Works with OpenAI, Anthropic Claude, Google Gemini, Qwen, or a local Ollama model. See Choosing a provider.

• Local or HPC execution. Run on your machine, or offload to a remote workstation or Slurm cluster with a one-line profile. LatticeMind never sends your API key to the remote host. See Remote & HPC execution.

• Multilingual. Prompts in English, French, or Chinese are understood and reports can be produced in the language you asked in.

• Self-describing projects. Every project is a self-contained folder of inputs, results, plots, status files, an event log, and a report — portable and reproducible.

Who this manual is for

This manual is written for users — materials scientists, chemists, and engineers who want results, not build systems. You do not need to know RESCU’s input syntax or how to compile anything. If you can describe a calculation, you can run it.

Ready? Continue to Installation to get LatticeMind installed, then to Getting started for your first project.