A partnership between the Department of the Air Force and MIT recognized staff on collaborative teams advancing artificial intelligence technologies for weather- and climate-related decision-making and optimized aircrew scheduling.

Two teams involving MIT Lincoln Laboratory staff were recognized with 2020 awards from the Department of the Air Force (DAF)−MIT Artificial Intelligence (AI) Accelerator. Recipients were selected from more than 150 airmen, researchers, faculty, and students from the DAF, Lincoln Laboratory, and MIT who are part of a collaborative effort to accelerate fundamental advances in AI. The AI Accelerator (AIA) was established in response to a 2019 White House directive to the Department of Defense to increase its focus on AI — an emerging field important to U.S. economic and national security. Research under the AIA seeks to harness the ability of machines to perform tasks normally requiring human intelligence, with the goal of improving U.S. Air Force (USAF) operations and addressing broader societal needs. All projects under the AIA develop “challenges” as a way to engage academia, industry, and the public in solving problems to advance AI. These challenges commonly involve open competitions to develop the best machine learning (ML) models to learn from or analyze publicly released datasets.

Building weather and climate resiliency

The AIA Director’s Award, which highlights “excellence and impact with a focus on collaboration across the AIA and with stakeholders,” recognized the Earth Intelligence Engine (EIE) project team for “cross-organizational collaboration, curation of novel datasets, visualization of forecasts, and delivery of an innovative challenge problem.” The EIE project targets three research areas — the earth intelligence platform, earth intelligence enhancement, and earth visual models — to build weather and climate resiliency for the USAF. Resiliency against hurricanes, wildfires, flooding, sea level rise, and other extreme-weather and climate-related threats is critical to protecting USAF resources such as military bases and ensuring mission readiness.

For the earth intelligence platform, the team is focused on making complex datasets on earth systems – the interacting systems of land, water, air, and life that make up the earth — AI ready. These datasets include atmospheric, terrestrial, and oceanic data from remote sensors like satellites, in situ sensors like aircraft, and numerical weather and climate models.

“A challenge in ML and AI in general is organizing and formatting the data so that they can be used to build and train ML algorithms,” said Lincoln Laboratory team lead Mark Veillette of the Air Traffic Control Systems Group. The other Laboratory team members are Allison Chang of the Air Traffic Control Systems Group and Esther Wolf of the Human Health and Performance Systems Group

A photo of Mark Veillette.
Mark Veillette

The team built an AI-ready dataset called Storm Event ImageRy (SEVIR), which they used to create an AIA challenge focused on short-term weather forecasting, or “nowcasting.” The SEVIR dataset contains images of thousands of weather events captured by satellite and radar. Competitors were challenged to generate future images based on these previous events as input. Nowcasts are used in many fields, such as air traffic control, humanitarian assistance and disaster relief, maritime operations, and public safety.

For the earth intelligence enhancement component, the team is researching ways to improve weather and climate forecasts by incorporating AI into the physical models of processes on which they are based, such as cloud formation and atmospheric turbulence. Weather and climate models driven by AI could accelerate their runtime, leading to quicker forecasting.

Ultimately, for a model’s outputted forecast information to be actionable, it must be presented in a visual, interpretable format.

“Often, meteorologists need to translate the information produced by the models for decision makers to actually use it,” explains Veillette.

To this end, the team is creating new visualizations of weather and climate impact to improve decision making. For example, the team applied AI to combine storm surge models with overhead satellite imagery to generate pictures of what floods will look like before they happen.

Actual and generated flood images.
The EIE team combined overhead satellite imagery of pre-flood areas with AI-incorporated physical models based on a generative adversarial network (GAN) — a class of machine learning framework — to predict what floods will look like.

In addition to targeting these research areas, the EIE team is engaging USAF stakeholders to educate them on the potential of AI in weather and climate forecasting. In the coming year, they will pursue opportunities to transition results of their research and challenge problem into USAF operational platforms.

The other members of the EIE team who contributed to the work recognized by this award are Brandon Leshchinskiy and Björn Lütjens of MIT and Maj. John Radovan of the USAF. 

Optimizing aircrew schedules

The AIA Challenge Award recognized the Puckboard project team for designing and implementing two community challenge problems focused on aircrew scheduling. Developed in 2019, Puckboard is a web-based software application for scheduling pilots and loadmasters (personnel responsible for loading and unloading cargo and passengers) to mission and training flights. Scheduling involves determining who is available to fly given other time commitments like deployments and vacation and who is qualified to fly based on their certification level and training status.

Though Puckboard was an improvement to the previous approach of planning squadron schedules on whiteboards and spreadsheets, it still required manually blocking out event times, ensuring crews have the right blend of qualifications for each event, and contacting airmen to assess their availability. Any change in plans — say, someone calling out sick or a higher-priority mission coming in from headquarters — would throw the entire schedule off, requiring a scheduler to rework it by hand.

Through the AIA, the Puckboard team has been working to incorporate AI into the software to automate this complex, time-consuming, manual process. The Lincoln Laboratory team members are Michael Snyder and Jessamyn Liu of the AI Software Architectures and Algorithms Group, Amy Alexander and Kendrick Cancio of the Air Traffic Control Systems Group, and Jeremy Kepner of the Supercomputing Center.

A photo of Michael Snyder
Michael Snyder

“We’re developing algorithms to make Puckboard more optimized to account for the various factors that go into scheduling and robust to last-minute changes,” says Snyder, who serves as team lead on the Lincoln Laboratory side. “Our algorithmic approach fuses two optimization techniques: linear programming and reinforcement learning.” 

To further improve Puckboard, the team partnered with the DAF Chief Data Office to roll out the two AIA challenge problems at virtual “datathons” held in July 2020 and April 2021. More than 100 airmen and other active-duty and reserves personnel participated in the datathons, events aimed at solving problems by using data. At the first event, participants were presented with anonymized historical scheduling data for C-17 aircrews and explored schedule optimization and visualization techniques. For example, one group introduced the concept of scheduling with some personnel and events locked in to account for instances when schedulers wouldn’t want to make changes because they would be too expensive. The Puckboard team has since incorporated this functionality into their algorithms. At the second event, participants developed techniques to parse through and associate data to link records of completed events back to their planned calendar entries. Linking this information is key to determining whether personnel were optimally assigned to potential events available to them.

A photo of the inside of a C-17 aircraft.
The team explores C-17 staffing firsthand.

“The value of these challenges is that they bring in individuals who weren’t already working in this problem space, so they have unique perspectives and approaches,” says Snyder. “Our team walked away from the events with lots of great insights, which we’re starting to weave into mathematical formulations and constraints for our algorithms. For example, schedulers want to know why a particular scheduling solution has been suggested, so we’re adding metrics on availability of personnel and ways to enforce the criticality of assignments.”

On a big-picture level, the team envisions the AI-assisted Puckboard could eventually be used not only to assign crew to events but also to reserve airspace, coordinate aircraft maintenance and repairs, and track personnel-qualification progress. The scheduling capability may also find application for unmanned aerial vehicles and fighter jets, and even outside the Department of Defense for civilian tasks, such as distributing medical-shift work and routing packages.  

The other members of the Puckboard team who contributed to the work recognized by this award are Capt. Ronisha Carter, Maj. David Jacobs, 2nd Lt. Luke Kenworthy, 1st Lt. Matthew Koch, Capt. Kyle McAlpin, Maj. Travis Smith, and Tech. Sgt. Allan Vanterpool of the USAF and Christopher Chin of MIT.

The DAF-MIT AIA also bestowed awards for scientific excellence and distinguished contributions. More information on these awardees is available in the official press release.