President Trump’s Golden Dome initiative has certainly created buzz within the Space industry. There are a lot of questions still unanswered, namely what the architecture of the system will be. Despite this, some inferences can be made as to what this space-based system will eventually consist of.
Systems already in the making, such as the Tracking Layer and Next-Gen OPIR will assuredly be part of the system; however, more will be needed if Golden Dome is to stop ICBMs and hypersonics. Space-based interceptors will likely be needed if boost-phase intercept is utilized within this system.
Whatever the architecture, the decision-making aspect may rely highly on Artificial Intelligence to make decisions. If the main thrust of the system is indeed boost-phase intercept, that gives the warfighter about three minutes to make a decision and then intercept, not a long time. AI will assuredly need to be integrated to some degree to maximize chance of a successful mission.
Breakdown of the Problem
The use of AI for boost-phase intercept is not a clearly defined application. Given uncertainty about the ultimate architecture of the Golden Dome, any account is therefore somewhat speculative.
However, some things can be noted up-front. AI models applied to threat-identification during the boost-phase of a missile launch must accomplish, at least, the accurate identification (based on incoming data) of a missile launch. Extended uses may include determining what kind of missile has been launched (ballistic, hypersonic, etc.) and re-formatting that information for ground systems or warfighters.
These tasks require that certain things be true of the model prior to such uses:
(1) The model must be able to interface with other components of the broader system in processing incoming data;
(2) It must appropriately identify a missile threat within a few minutes;
(3) It must have a means of rating its confidence in the identification (say, a simple percentage) as well as back-ups and fail-safes;
(4) Finally, there must be adequate underlying hardware for rapid data-processing.
Is AI Up to the Challenge?
Concerning (1), the most fruitful integration of AI may be either within the Space Force’s Overhead Persistent Infrared program for warning and tracking missile threats or within the low-Earth orbit Tracking Layer that provides global and persistent missile warning and tracking (or both). The latter in particular is comprised of a network of satellites equipped with infrared (IR) sensors to detect and track heat signatures from missile threats.
Sensor fusion research, some of which is carried out by the U.S. Air Force Research Laboratory, is already experimenting with machine learning (ML) for similar needs. The basic idea is to relieve a data-processing bottleneck by integrating sensors with varying but complementary strengths and weaknesses and using ML to reformat the collected data into standardized formats, making them human-usable. The upshot is that such sensor fusion delivers on the promise of rapid data processing while improving the robustness of the data analysis in challenging environments.
The application to Golden Dome’s dispersion of space-based interceptors is a different beast, to be sure, but data collected by IR sensors would need to be adequately processed before a threat is identified, let alone further analyzed and translated for human-use. This would require interfacing between the models and the sensors (potentially aboard several different platforms), analyzing it, and potentially re-formatting it before eventual transmission to warfighters.
Even before this, however, the problems of accuracy and reliability in (2) and (3) loom large. An ML model requires training on relevant data so it can learn to make the appropriate classifications of input data. Such data, if it exists, is largely classified. However, whether the proper kinds and amounts of data exist, classified or not, is not at all clear.
Synthetic data – having the model generate data based on some kind of human-given procedure – could play a role here, but such techniques tend to be brittle when moved into “open” domains (as opposed to “closed” domains where all the relevant information is contained within the problem it faces). Synthetic data is a gamble.
Thus, significant concerns about the accuracy and reliability of resulting outputs from these ML models in production should be raised, as it is not clear how accurate the identifications made by this system would be.
If a machine learning model is processing data collected by IR sensors, the only outputs that model can generate will be derived from the relationships between the data they trained on; this can yield impressive results, but it is not a precision exercise.
Some note the potential for strategic ambiguity here – an adversary does not know when the models have failed to make the correct identification – but ML models often share common pitfalls no matter their application; such ambiguity may be therefore be tested by adversaries who understand these pitfalls.
This is why AI is typically best suited for “closed” domains where the available information is all the model needs, and where there is room for error (a point made generally and correctly by Shield AI’s Ryan Tseng). Space-based interception of a missile launch during its boost phase is very much an “open” domain where the available information is incomplete and room for error is exceedingly low given the time-constraints.
It is also correct to note, as the Center for Arms Control and Non-Proliferation recently does, that several advancements in “sensor coverage, battle management and interceptor reliability, not to mention substantial new infrastructure investment,” are required for Golden Dome.
AI is therefore best situated as a set of modules within a broader, much more complex system which does not rely solely on their outputs, as they could represent a single-points-of-failure. (And other technologies are likely better suited for things like trajectory calculations, once identification of a missile launch and type is confirmed.)
Finally, the problem in (4) is that the underlying hardware must be sufficient to handle rapid data-processing and communication with other aspects of the Golden Dome. Sensor fusion research provides another precedent here: edge computing—the use of the platform to which an AI model is equipped as a mobile data center, so to speak—is gaining traction. The extent to which scalability of the overall architecture remains a challenge, however, and should not be underestimated.
There are still a lot of questions surrounding the Golden Dome initiative. With time, more information will become available that will answer some of the question however, not all. If boost-phase intercept is one of the pillars of Golden Dome, AI will play a part.
Industry Analyst
Education: MA, Political Science, Villanova University; BA, Political Science & Philosophy, Rosemont College.
Vincent Carchidi has a background in defense and policy analysis, specializing in critical and emerging technologies. He is currently a Defense Industry Analyst with Forecast International.
Carter Palmer has long held a keen interest in military matters and aviation. As a FI's space systems analyst he is responsible for updating the reports and analyses within the Space Systems Forecast – Launch Vehicles & Manned Platforms and Space Systems Forecast – Satellites & Spacecraft products.
