MIT Sustainable Design Lab
Accelerating Urban Energy Modeling
MIT Sustainable Design Labs’ primary challenge was to build an application that can be continuously run by urban policymakers and energy modelers to evaluate carbon reduction potential in the building stock as they race to meet emissions goals. The team dealt with long training loops while prototyping model architectures to optimally train on 1.5MM – 2MM sample data sets that took approximately 30 hours per epoch.
The MIT team was slowed down by writing and debugging noisy boilerplate code in their model, which suffered from a general lack of readability. Iteration speed was also a great challenge. Junior team members were blocked when onboarding and first setting up their environments for weeks. Even senior researchers felt friction when bringing their individual prototypes to production, reducing overall team productivity. Iteration speed was extremely important because of the academic context where experimentation with new datasets and architectures is always underway.
“”Our training loops went down from 36 hours to 6 hours and Studios increased productivity by 33%, from 3 weeks to 2 weeks per iteration.”
Using Lightning Studios’ environment management increased productivity by 66%, instantly. The SDL team was also able to use WandB for logging and model registration without the challenges of integrating a new API into the codebase because Lightning Trainer automatically configured everything needed.
Additionally, adopting Lightning Trainer allowed the team to organize their training code and eliminate hard-to-read boilerplate code. This was essential in speeding up iteration time between training loops and allowed the team to experiment and switch between different datasets and model candidates. They were also able to speed up training by using the PyTorch Lightning OSS framework to organize their boilerplate and Studios to get powerful multi-GPU machines that they didn’t previously have access to.
- Reduce model training time by 85%, from 36 hours to 6 hours
- Reduce prototyping and experimentation time by 60%, from 5 weeks to 2 weeks, by allowing for efficient collaboration between senior and junior researchers
- Reduce time to take model from prototype to production by 33%, from 3 weeks to 2 weeks
- OSS framework improved the organization of the model code and make it self-documenting, improving readability and reusability
- Reducing onboarding time for new research assistants by 90%, from 2 weeks to 2 days by eliminating environment management and enabling them with powerful GPUs
About the MIT Sustainable Design Lab
The MIT Sustainable Design Lab, operating under the School of Architecture and Planning, spearheads initiatives to cut down the carbon footprint of buildings, which are responsible for 40% of global emissions. Led by Christoph Reinhart, the lab developed UBEM.io, an app for evaluating energy retrofit strategies, widely utilized by cities and policymakers for project modeling. Sam Wolk, the Lead Developer UBEM.io, led the migration from distributed tools to Lightning AI as they bring their model to production.
MIT Sustainable Design Lab is set to deploy its model as a web API and will continue refining simulations with Lightning Studio. Their commitment to rapid iteration and simulation acceleration remains in lockstep with Lightning AI’s capabilities.