Discussion

Points that were raised during the writing of this document in discussions between participants are stored here:

Discussions


Roadmap

The objective of this project is to create an interoperable ecosystem compatible with various applications, primarily focused on differentiable physical modeling. Our approach places specific emphasis on neural network potentials (NNPs) and Free Energy calculations. Our architecture is designed with clean APIs that provide developers with simplicity and minimal overhead requirements, meaning that it will require less additional resources and time to implement and balances efficiency, versatility, and functionality. All packages will be designed with interoperability in mind, ensuring that they can be used in conjunction with one another, as well as with other systems, to maximize versatility and functionality.

Overview

We plan to offer three distinct, interoperable, and modular packages designed to function in synergy:

Differentiable MD Package and Free Energy Package (package name: chiron): This component will focus on general capabilities for Markov Chain Monte Carlo (MCMC) based simulations and sampling algorithms. Additionally, the package will include features for calculating free energies, sampling protonation states, tautomeric states, and other useful application algorithms built on top of these capabilities, such as replica exchange free energy calculations. The primary focus will be on performing these calculations using Neural Network Potentials (NNPs), allowing for efficient and accurate modeling of complex molecular systems. The technical roadmap is outlined here:

<aside> 💡 Technical Roadmap Chiron

Chiron Technical Roadmap

</aside>

Neural Network Package with Training Routines and Datasets (package name: modelforge): This package will be centered around neural networks, including essential training routines to create, optimize, and store models effectively. Datasets will be provided to enable accurate training and validation of the neural network structures. The technical roadmap for the modelforge package is outlined here:

<aside> 💡

Technical Roadmap Modelforge

Modelforge Technical Roadmap

</aside>

Benchmark/testing Package/Website (package name: auditorium): This package/website will be focused on performing the training of the NNPs, selecting appropriate hyper-parameters, performing benchmark and stability tests. We want to upload and display these results to a website so that every user can decide which NNP best fits the need of a downstream. OGB is a good example of what we want to provide (http://ogb.stanford.edu). The website that displays the training, benchmark and stability results can be as simple as a github page.

<aside> 💡 Technical Roadmap Auditorium

Auditorium Technical Roadmap

</aside>

User Requirements and Target Audience

There are general objectives that we need to meet so that user feel confident in using this package: