AMMap

Additive Manufacturing Mapping of Compositional Spaces with Thermodynamic, Analytical, and Artificial Intelligence Models

AMMap is a research-grade tool utilizes our (a) new algorithm for representing complex compositional design spaces in AM as graphs, generated by high-performance nimplex library discussed in our npj Unconventional Computing paper (10.1038/s44335-024-00012-2), (b) established algorithms for optimal graph traversal, (c) thermodynamic models for phase stability and solidification, and (d) other scientific property models; in order to explore and navigate possible alloys which can be made, well beyond what can be visualized by humans, given constraints like using up to 4 feedstocks at the time out of 12 possibilities to choose from. Please refer to our recently-accepted JPhysMat paper on AMMap itself (10.1088/2515-7639/ade73e) for full description.

Furthermore, AMMap generates reusable callables (with easy integration into AI pipelines) for several different CALPHAD based methods out-of-the-box and is set to grow rapidly in the near future. These include thermodynamic equlibrium calculations with pycalphad, Scheil-Gulliver solidification with scheil, and 5 different models for predicting cracking susceptibility discussed in this publication.

In the current "default" implementation, results coming from all these methods are used to establish multi-constraint feasible subspace by exploring the nimplex graph representation of the design space without going into infeasible regions and find shortest/optimal paths between N desired alloys using graph-based path planning tool of choice.

We have set up a cloud-based GitHub Codespace can be used for all Jupyter notebook exercises for free with just a GitHub account. Your progress will be persisted for some time (depending on your settings) so it can be used for research activities too; however, the free version may lack the computational power to perform some heavier calculations (e.g., Scheil-Gulliver) fast; thus, we recommend that power users set it up locally or an HPC node, evaluating notebooks with a tool like papermill, in the background, or fetching pre-computed results (see scheilmap.json) from scripts to then be analysed in Codespaces.

Quick Start:

Capabilities

Note Full technical discussions is given in our recent JPhysMat paper (10.1088/2515-7639/ade73e). Section below is a short highlights of tool ability.

1. Simplex/Compositional Graph Generation to allow for compositionally complex materials with high order of possible combinations to be fully considered in a homogenous fashion.
2. Stitching of elemental spaces from different thermodynamic databases into singular traversable graph to allow for path planning across multiple composition regions from incompatible databases due to model differences.
3. Material Information Generation Thermodynamic (Equilibrium and Scheil-Gulliver) information about the material at any given graph point calculated to determine phase composition to avoid undesired phase formation in path planning. Thermodynamic information can then be used to determine the hot-cracking susceptibility of a point to further inform material design. Visual representations of capabilities for example systems can be seen at end of README below.
4. Infeasibility Gliding Detects infeasible regions and avoids uneccessary calculations of interior points to reduce total computational cost.
5. Path Planning Compatibility Deployment of any graph algorithms on created design space and highly compatible with backends of choice. Currently finds the shortest path which can be stretched. Found path can then be simplified to change from point-to-point path to generalized format.

• In simple cases, like path planning across 3 ternary spaces using on 3 different thermodynamic databases, to go from from approximated SS304L to pure titanium, the results generally look like this:

• In more advanced cases, the results can no longer be visualized in 2D or 3D exactly, as graphs start to be hyperdimensional (5 and more components) or intersect each other, like in the example below from our paper. This, in no way, affects the computation process, which can produce results for, e.g. all 6-components spaces in 12 component space given enough compute power, but to visualize the results we have to rely on either (a) relaxed embeddings based on graph connectivity with some "energy" models,

or unfolding the lower-dimensional (3/4 component) spaces of interest for analysis and planning,

• See a couple of additional visualizations in the Examples section.

Planned

1. YAML format input for human readable definition of design space, constraints, and objectives
2. Expanded Property Models More material properties that can be mapped onto compositional graphs are planned, e.g. elastic modulii
3. Output of callables for Machine Learning
4. Path Planning Updates to minimize the number of turns or more generally hyperplanes.
5. Build Planning Algorithm Collaboration to create algorithm to convert given path into physical AM build path for fabrication

Installation

Conda Environment and Installs

It is recommended to use new environements for all python projects, this can be done as follows:

conda create -n AMMAP python=3.11 liblapack jupyter numpy pandas plotly scikit-learn

conda activate AMMAP

Or, if your environment already exists, simply:

conda install -y python=3.11 liblapack jupyter numpy pandas plotly scikit-learn

Clone repository

Clone the github repository in order to have the jupyter notebook and callables

git clone https://github.com/PhasesResearchLab/AMMap.git

nimplex

The primary installation requirement is nimplex, which requires the small and easy-to-install Nim (Installation Instructions) compiler (assuming you already have a C compiler), which can be done with a single command on most Unix (Linux/MacOS) systems:

• using conda cross-platform package manager:

  conda install -c conda-forge nim

• on MacOS, assuming you have Homebrew installed:

  brew install nim

• with your Linux distribution's package manager (note that it may be an outdated nim version, impacting performance), for instance on Ubuntu/Debian Linux:

  apt-get install nim

Then, you can use the boundeled Nimble tool (pip-like package manager for Nim) to install two top-level dependencies: arraymancer, which is a powerful N-dimensional array library, and nimpy which helps with the Python bindings. You can do it with a single command:

nimble install -y arraymancer nimpy

Now, you can update the nimplex submodule repository and compile it for AMMap. You want to do so in a way that creates its Python bindings. You will need slightly different flags depending on your system configuration, but for Unix (Linux/MacOS) you can do so with commands below, after making sure you are in the root AMMap directory:

git submodule update --init --recursive
nim c --d:release --threads:on --app:lib --out:nimplex.so nimplex/src/nimplex.nim
nim c --d:release --threads:on --app:lib --out:utils/plotting.so nimplex/src/nimplex/utils/plotting.nim

For Windows and other platforms, you should consult nimpy instructions.

CALPHAD Tools

When you are done, you should also install pycalphad and a forked version of a python package for scheil found here

pip install git+https://github.com/HUISUN24/scheil.git
pip install pycalphad

Optional Pathfinding used in Example

pip install pqam-rmsadtandoc2023 pathfinding

Other useful packages

To run Jupyter notebooks from the command line, especially useful on an HPC, you should install papermill:

pip install papermill

Example Outputs

Equilibrium

Cracking Criteria

Path Planning with Infeasibility Gliding

Contributing Guidelines

AMMap is an MIT-licensed open-source project, and we welcome contributions from the community. As of June 2025, it is still a research-grade tool, but we are actively working on adding high-level of automation and other usable features to turn it into a user-tool that can be used by additive manufacturing researchers and engineers without significant programming experience. The goal is to have a simple YAML-based input format to describe the design space and constraints, and then use the tool will run everything automatically on your machine or on an HPC.

If you are interested in contributing to the core functionalities, please reach out to us via GitHub Issues and we will be happy to discuss your ideas and help you get started.

If you are interested in contributing scientific / ML / thermodynamic models to our system, that can already be done in a relatively straightforward fashion, by providing a callable template script (see existing ones). Feel free to do a pull request directly but do not hesitate to reach out if you have any questions or need guidance.

Please Cite

If you use AMMap in your research in any capacity, please cite the following two papers:

• Krajewski, A.M., Beese, A.M., Reinhart, W.F., Efficient generation of grids and traversal graphs in compositional spaces towards exploration and path planning. npj Unconv. Comput. 1, 12 (2024). DOI: https://doi.org/10.1038/s44335-024-00012-2

• Richter, A., Krajewski, A.M., Yang, Z., Beese, A.M., Liu, Z.K., AMMap Tool for Additive Manufacturing Design, Alloy Discovery, and Path Planning. Journal of Physics: Materials, Accepted (2025). DOI: https://doi.org/10.1088/2515-7639/ade73e

If you use the thermodynamic models through pycalphad, please also cite the following papers:

• Otis, R. & Liu, Z.-K., (2017). pycalphad: CALPHAD-based Computational Thermodynamics in Python. Journal of Open Research Software. 5(1), p.1. DOI: http://doi.org/10.5334/jors.140

If you use the cracking susceptibility models, please also cite the following paper:

• Yang Z., Sun H., Liu Z.-K., Beese A.M., Design methodology for functionally graded materials: Framework for considering cracking, Additive Manufacturing, Vol 73, 2023, 103672. DOI: https://doi.org/10.1016/j.addma.2023.103672