Learning Astrophysical Uncertainties in Dark Matter Direct Detection with Simulation-Based Inference

Deep in the universe, most of the matter is not the stuff we can see, touch, or measure with ordinary tools. It is invisible, known only by the way its gravity pulls on stars and galaxies. Scientists call this dark matter, and while we know it makes up about a quarter of everything in existence, we have no idea what it is made of. One leading idea suggests it consists of heavy, slow-moving particles that rarely bump into normal matter. If these particles exist, they should occasionally strike the nuclei of atoms inside a detector on Earth, creating a tiny flash of light and a small electrical signal. Finding this faint signal would be a monumental discovery, revealing the true nature of the dark universe.

The challenge is that these signals are incredibly rare and easily confused with background noise from natural radioactivity or other cosmic rays. To find them, scientists use massive tanks of liquid xenon, chilled to near absolute zero, buried deep underground to shield them from cosmic interference. When a dark matter particle hits a xenon atom, it produces two distinct signals: a flash of light and a burst of electrons. By measuring the ratio between these two signals, researchers can tell if the event was caused by a dark matter particle or by a common background particle. However, the exact pattern of these signals depends on how fast the dark matter particles are moving as they drift through our galaxy. This speed is not known with perfect certainty, and different assumptions about the galaxy's structure can change what the signal looks like.

In a new study, researchers Felix Kahlhoefer and Niklas Reus from the Karlsruhe Institute of Technology in Germany have developed a clever way to handle this uncertainty. Instead of trying to calculate complex mathematical formulas to account for every possible variation in the galaxy's structure, they used a method called simulation-based inference. They built a computer program that learned to recognize the difference between a dark matter signal and background noise by studying millions of simulated examples. In these simulations, the computer was shown data generated under many different assumptions about how fast the dark matter particles were moving. By training on this wide variety of scenarios, the computer learned to ignore the specific details of the galaxy's motion and focus only on the core features that indicate a dark matter hit.

The researchers tested this approach using a model of the XENONnT experiment, one of the most sensitive dark matter detectors currently operating. They created a digital twin of the detector, simulating how it would respond to dark matter particles of different masses and strengths of interaction. They then trained a neural network, a type of artificial intelligence, to act as a classifier. This network was shown pairs of data: one set where the dark matter signal was present, and another where it was absent. The goal was for the network to learn a rule that could distinguish the two, effectively calculating the likelihood that a specific set of observations came from a dark matter source rather than random noise.

What makes this method particularly powerful is how it deals with the unknown speed of the dark matter particles. In traditional analysis, scientists often have to fix the speed of these particles to a single, standard value, which can lead to errors if that value is wrong. In this new approach, the researchers simply included many different speed models in the training data. The computer learned to treat the speed as a variable that could be anything within a reasonable range. As a result, when the network makes a prediction, it has already accounted for all those different possibilities. The uncertainty about the galaxy's structure is effectively averaged out, leaving a clear answer about the dark matter itself without needing to do extra calculations for every new piece of data.

The study showed that this method works remarkably well. When the researchers tested the trained network on simulated data, it produced results that matched the expectations of the most established statistical methods used in the field. The network successfully identified the correct mass and interaction strength of the dark matter particles in the simulations, even when the data was noisy or when the underlying assumptions about the galaxy were very different from the standard model. In fact, the network performed so well that it could handle extreme scenarios where the standard assumptions were completely wrong, still producing reliable results. This suggests that the method is robust and can adapt to the messy reality of the universe better than older techniques.

Perhaps most importantly, the researchers demonstrated that this approach can be scaled up easily. As experiments become more sensitive and probe deeper into the "neutrino fog"—a background of signals from solar neutrinos that will eventually obscure the dark matter signal—the need to account for more complex uncertainties will grow. Traditional methods would struggle with this, requiring massive amounts of computing power to re-run calculations for every new variable. The simulation-based method, however, handles this complexity during the training phase. Once the network is trained, it can analyze new data instantly, regardless of how many different astrophysical models were included in the training. This makes it a practical and efficient tool for the next generation of dark matter searches.

The findings confirm that artificial intelligence can be a vital partner in the hunt for dark matter, not by replacing human understanding, but by handling the heavy lifting of complex statistics. By learning from a vast library of simulated universes, the computer can find the signal in the noise with a clarity that matches the best human-designed methods. This does not mean dark matter has been found; no such signal has been observed yet. But it does mean that when the next experiment turns on, scientists will have a sharper, more flexible tool to interpret the data, ready to distinguish a true discovery from a trick of the light, even if the galaxy behaves in ways we did not expect.