\(\renewcommand{\AA}{\text{Å}}\)

4. Information for Developers

This section describes the internal structure and basic algorithms of the LAMMPS code. This is a work in progress and additional information will be added incrementally depending on availability of time and requests from the LAMMPS user community.

A discussion of software engineering methods applied to LAMMPS over time and a general outline of the design and maintenance approach by the LAMMPS developers can be found in the paper LAMMPS: A Case Study For Applying Modern Software Engineering to an Established Research Software Package, in the USRSE’25 conference proceedings.

On using AI coding agents

Recently AI coding agents have seen massive improvements and thus have become quite popular and viable in assisting with the development of new code or with adding features to existing code. The LAMMPS developers have made some experiments and tried to assess the how effective and reliable the results are. Below are some of our observations.

A significant part of LAMMPS was written by taking some existing classes that perform similar calculations or operations and then replacing the functionality. That kind of task is well suited for a coding agent, but there are often little details that require some manual adjustments and thus careful checking of the resulting code is recommended. Similarly, porting styles to the KOKKOS or OPENMP package can be delegated in many cases, provided suitable template examples exist and the coding agent is asked to look at those. Github Copilot can also run LAMMPS unit tests and even add unit tests for new styles if a suitable template already exists. Asking a coding agent to create new functionality for which there is no similar code in either the LAMMPS sources or some other repository has a much lower chance of success.

In general, it is best to be explicit and verbose when assigning tasks to coding agents. It can also helpful to ask the agent to break up complex tasks and create a master plan. Avoid asking Github Copilot to do too much at once. For example, we have found that porting 5 or so styles to the KOKKOS or OPENMP package is reasonably efficient, without running into session timeouts. When a session is finished, asking the agent to summarize what it learned and including that context in the next session can also improve efficiency.

Another task that can often be passed successfully to a coding agent is the creation of suitable documentation. However, the results have a tendency to be overly verbose and sometimes redundant, so some manual editing to tighten it up is recommended.

Finally, coding agents are quite good at reviewing existing code. The LAMMPS developers use GitHub Copilot regularly to review submitted pull requests and over time the rate of incorrect advice has become rather low. When using an AI coding agent, it is generally recommended to use a different model for the code review than what was used to write the code.

To assist coding agents in their tasks, and particularly to provide instructions that GitHub Copilot will automatically read, we have created a file .github/copilot-instructions.md that can be used to inform coding agents about important steps when creating or reviewing code for LAMMPS.

When contributing code to LAMMPS that was at least partially created by a coding agent, please provide information in the pull request description to the extent of AI usage and which model or models were used.