The following blog post is adapted from a research paper presented at the Naval Postgraduate School’s Annual Acquisition Research Symposium. You can download the paper from the Navy Postgraduate School here.
As competition between the United States and near-peer adversaries heats up, the U.S. Navy faces increasing challenges to its sea dominance, particularly in contested regions like the South China Sea. China’s behavior has grown increasingly aggressive in the maritime domain, from using water cannons or lasers on Philippines resupply vessels, to building a constellation of military bases in contested areas, among other concerning incidents.
To credibly project power in the region, the United States needs superior Naval capabilities and platforms that deter sophisticated adversaries. They must be online, available, and ready to be called upon if needed.
Unfortunately, that availability hinges on the U.S. industrial base that produces the parts that keep platforms running. Right now, the industrial base does not have the speed and capacity required to swiftly address operational challenges. The Department of Defense has initiated large-scale strategic change to strengthen its industrial base in the long-term. However, obsolescence issues such as parts shortages and dwindling supplier counts still plague weapons systems, creating risks to fleet readiness.
Luckily, artificial intelligence and machine learning technologies can solve this challenge at speed and scale. The United States can leverage them right now to weed out risky parts and identify alternatives before any disruption happens.
Currently, the process to identify alternate parts is time- and labor-intensive. Once a part is deemed unable to meet fleet requirements (or at risk of such), Navy engineers are notified. Then, they scour hundreds of sources by hand, comparing a multitude of technical characteristics, to find alternatives. Beyond its time-consuming nature, this process does not always yield fruitful results. The myriad data sources mean that potential replacement solutions can be overlooked.
To accelerate this process, Govini developed a scalable, technology-forward methodology: use an LLM to analyze the potential form, fit, and function of current parts at scale, identify potential alternates, and pinpoint the ones with stock that’s already on-hand.
What does this look like in practice? Govini deployed this LLM-based AI model in our Ark software platform, supporting the Ohio-Class Submarine program office to accelerate their alternate parts process.
Preparing the Data and Training the Model
Govini turned to its National Security Knowledge Graph (NSKG) in Ark to build a baseline dataset for analysis, bringing together all of the relevant data that could possibly describe the form, fit, and function of parts for a selected weapons system. The National Security Knowledge Graph is powered by Govini’s Object Fusion data engine, which continuously ingests, normalizes, and integrates new data sources with existing data catalogs.
Using input from subject matter experts, Govini then trained the LLM model, ensuring that it assessed parts across specific characteristics like a part’s weight, material, size, and description.
With the baseline dataset and model ready, it was time to begin. The dataset of the Ohio-Class Submarine parts was evaluated to identify the parts with the longest lead times–the ones which would cause severe disruption if they needed to be replaced fast. Those parts were run through the LLM model to identify similar ones where stock was already on-hand, and then rank those options on a similarity score, out of 100. Any identified alternates with a score over 90 were considered strong candidates to replace the risky part.
AI for Improved Readiness
Drawing on the machine speed of artificial intelligence, the model scanned the approximately 123,564 unique parts associated with the Ohio-Class Submarine, as identified by 25 different Weapons Systems Designator Codes defined by the Defense Logistics Agency.
Some of those parts had lead times as high as 1,261 days. It could take as long as 42 months to get the part to the submarine and get the sub repaired, with potentially drastic consequences for fleet readiness.
So what options did the program office have? Running Ohio-Class Submarine parts through the LLM model spotlighted the 68,569 unique parts that had alternate options with high similarity scores, over 90. To further refine the list, the model filtered it down to remove parts where there was no available stock on hand. This surfaced the alternates where there was a quick solution. Ultimately, the AI model pointed to a number of available alternatives for vital parts.
Zooming in on one of them provides a case in point.
NIIN 00501959 is an Annular Ball Bearing, part of an auxiliary system on the Ohio-Class Submarine. It has a lead time of over 1,000 days–so without an alternative, waiting on this part to arrive would dramatically impact shipyard schedules. Govini’s LLM found four alternates with stock-on-hand. To verify results and put a human in the loop, Govini checked its work in a detailed, manual evaluation across 18 unique technical characteristics.
With this short list in hand, engineers working on the Ohio-Class Submarine have essentially already found the needle in the haystack, without having to dig for it.
Artificial intelligence and machine learning streamlined the painstaking process of comparing myriad data sources. Instead, the Ohio-Class Submarine team can leverage these results and quickly confirm that one of these similar parts can be used in place of the original Annular Ball Bearing.
The ability to quickly identify alternative parts enables the U.S. Navy to maintain shipyard availability schedules and overall fleet readiness more effectively. Leveraging AI and ML to analyze large data sets at scale will expedite a previously manual process. Tactically, utilizing these technologies will save the Navy time and man-hours, freeing up time to focus on other vital missions. Strategically, faster discovery of alternate parts can mitigate overall schedule impacts to the fleet. Artificial intelligence and machine learning are powerful shipmates.