LINK: A $1M DARPA Contract to Help Scientists Build Better Multiphysics Models

We’re pleased to announce Galois’s LINK project, part of DARPA’s Computable Models (COMPMods) program. The project aims to make it easier for scientists to build multiphysics models with an end-to-end, automatic, and portable solution. 

Multiphysics models are challenging to create, customize, and reuse. One problem is the software and hardware are often outdated, which makes it difficult to leverage solutions for more than one physics type. Even when legacy systems use one type of physics, the code combines the problem and solution, and this is not always transferable to a different kind of physics that needs to be modeled. 

For us, working with scientists and thinking of ways to solve their domain problems with custom tools is a dream come true. 

The LINK project team consists of computer scientists with strong engineering skills, so we have proactively engaged with domain experts during the development process to understand their workflow, current tools, and issues.  We have been working on designing an interface with those practicalities in mind. 

A Textbook Multiphysics Problem 

Let’s use NASA as an example. NASA is interested in designing vehicles with increased payloads suitable for launching advanced robotic and human missions to Mars. However, the payload requirements for these missions exceed current capabilities, which use traditional parachutes to decelerate.

One potential solution is reducing the vehicle’s speed for entry, descent, and landing with aerobraking and retro-propulsion. NASA wants to model how the thermal protection material is affected by the retro-propulsion system during descent. 

This is a “textbook” multiphysics problem with multiple components: fluid flow, thermodynamics, and solid mechanics. The problem is that there is a manual translation to high-performance code, models for each physics component are created independently, and they all use different solvers. There is no unified system to address the missing features.

To address this problem, we’re designing LINK to help domain scientists separate what they want to compute from how they want to compute it.

LINK would provide scientists with domain-specific languages (DSLs) that use terminologies familiar to each scientist’s domain, making it easier for scientists to quickly and accurately build models. LINK would have several DSLs, giving domain experts unprecedented freedom to prototype for particular types of physics. 

We want to empower scientists to define their problem and trust LINK to automate the rest. That includes writing a problem definition and coupling in DSLs, a compatibility check for the independent models, and generation of high-performance code.  

We believe LINK’s use of domain-specific tools to define domain problems will result in more reliable and efficient code. We also think it is a superior solution to the current scenario where domain experts work with general programming languages. LINK would provide clear abstractions for the users in their specific domains of interest.  

Our objectivity allows us to take a step back and consider making the whole modeling toolchain better and creating impactful change. 

To learn more about LINK, please see our project page

Approved for public release; distribution is unlimited. This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Agreement No. HR00112090064