The Air Force Institute of Technology accepted the challenge to educate data scientists within the Air and Space Forces along with the rest of the DoD.

The Department of Operational Sciences initiated an 18-month in-residence master of science in data science program with core courses providing mathematical and statistical foundations, along with education in data engineering, as well as knowledge in developing artificial intelligence and machine learning algorithms.

Program electives allow graduate students to explore diverse topics, such as computer vision, data wrangling, text mining, and reinforcement learning while the research thesis hones their skills to address real DoD challenges using and exploiting data for better performance.

The first cohort in the data science degree started in September of 2021 and will graduate in March of 2023. 

Beginning with the thesis research from its first cohort, AFIT is developing numerous lines of relevant research spanning data on medical, contracts, and target images. A partial list of examples of cutting edge research AFIT data science students are conducting include developing insight to tailor virtual reality pilot training to ensure trainees maintain optimal stress levels via psychometric data; creating new computer vision models which output accurate classifications with well calibrated certainty measurements; and enhancing algorithms to engage in autonomous air-to-air dogfighting.

The Operational Sciences Department plans to start a Ph.D. in Data Science beginning the fall of 2023. Potential Wright-Patterson AFB students (military, civilian, or contractor) should contact the department at AFITensDataAnalytics@afit.edu. 

Units desiring officers with data science degrees should code their billets with the academic code OCCE.

Organizations with significant data challenges can discuss potential research collaborations with Dr. Bruce Cox, program director for data science.
 

Data Science Successful Outcomes Identified Through Student Theses 

A Conformal Prediction Approach to Quantify Student Pilot Error via Multimodal Physiological Signals
By: 2d Lt. Gregory Barry
Sponsors: AFOSR and MIT Lincoln Labs
Using physiological and positional data recorded from simulated flights, we find that within this multimodal recording, there are key features which allow for accurate, autonomous scoring of pilot error. Through correctly predicting error in training flights within an allowable range, we can eliminate mundane duties of instructor pilots and focus them instead on personalized training for each student pilot. The methodological process we implemented throughout our research transforms vast physiological recordings of pilot data to implementable models which quantitatively define a student pilot’s performance. Lt. Barry presented this research to individuals from MIT Lincoln Laboratory and MIT in Boston, MA, on August 23, 2022. He also presented his research to the DAF-MIT AI Accelerator on November 4, 2022.

Generation of Beyond Visual Range Air Combat Tactics via Reinforcement Learning
By: 2d Lt. Caleb Taylor
Sponsor: Strategic Development Planning & Experimentation
A one-versus-one air combat problem is considered wherein a friendly autonomous aircraft must engage and defeat an adversary autonomous aircraft in a beyond visual range environment. Both autonomous aircraft employ medium-range air-to-air missiles and guns. The Advanced Framework for Simulation, Integration, and Modeling is leveraged to model the complex and interdependent operations of aircraft, sensors, and weapons. A reinforcement learning solution procedure that leverages a multi-layer neural network was developed and tested to determine effective maneuvering and weapons tactics. This research supports ongoing development of improved modeling, simulation, and analysis techniques for enhanced analysis of USAF capability development activities.

Supervised Multimodal Spatio-Temporal Smooth Sparse Decomposition for Lightning Prediction
By: 2d Lt. Grace Metzgar
Sponsor: SpOC/S9A
The uncertainty of lightning greatly impacts a variety of weather-sensitive operations inciting the development of techniques used to predict its behavior. However, the complexity of weather data increases the difficulty of creating models that can accurately recognize and forecast lightning patterns. To address these shortcomings, we developed supervised spatio-temporal smooth sparse decomposition (ST-SSD) to detect anomalies within high-dimensional streaming data to predict the time and location of future anomalies. We employed the Storm EVent ImagRy (SEVIR) dataset which was created to aid in the inefficiencies of developing these complex models by combining spatial and temporal weather data from multiple sensing modalities. We hypothesized that by applying supervised ST-SSD to such data will allow for greater accuracy in predicting the time and location of lightning strikes. Lt. Metzgar presented her research at INFORMS 2022.