At the heart of solid-state physics, a material's Fermi surface defines its fundamental electronic properties and, by extension, its behavior in applications ranging from superconductivity to next-generation electronics. For decades, researchers have relied on techniques such as angular correlation of electron-positron annihilation radiation (ACAR) to three-dimensionally reconstruct the two-photon momentum density (TPMD), a process that allows these surfaces to be visualized in detail. However, the road to a reliable rebuild is fraught with obstacles: high noise, acquisition times that can stretch for months, and the need for multiple angular projections. Now, a new family of algorithms dubbed DeepCormack promises to revolutionize this field by integrating artificial intelligence into the reconstructive process, offering not only a drastic improvement in image quality, but also a significant reduction in the time needed to obtain useful data.
The proposition behind DeepCormack is elegant and deeply practical. Rather than treating the reconstruction problem as a purely statistical black box, the researchers have opted for a hybrid approach: they combine Cormack's method—a classic adapted to crystal symmetry—with supervised deep-learning models, such as convolutional networks (CNNs), multilayer perceptrons (MLPs), and UNet architectures. The key is that these models are not trained with huge experimental databases – which are practically non-existent for this domain – but with realistic synthetic volumes generated from a single calculation of reference moment density obtained using density functional theory (DFT). The technique employs decomposition into singular values and decomposition into dynamic modes to produce a variety of training volumes that faithfully mimic the variations and noises of a real experiment.
From a technical perspective, the progress is substantial. In tests with synthetic data, DeepCormack achieves an improvement of approximately 8.5 dB in peak signal-to-noise ratio (PSNR) over the traditional method, even with counts of 200 million events. But most importantly, performance remains stable when counts are drastically reduced, suggesting that acquisition times could be shortened from weeks to days without sacrificing reconstructive quality. This has direct implications for materials research: a scientist who previously needed months to characterize a sample could now get results in a much more manageable timeframe, accelerating the discovery of new compounds with exotic electronic properties.
However, the method is not a universal magic solution. Its generalization to experimental data critically depends on whether the training distribution generated from the DFT calculation matches the actual material being measured. That's why the authors recommend pairing DeepCormack with a DFT calculation specific to the target material, thus creating a tailor-made training dataset. This need for customization fits perfectly with a broader trend in the world of software and technology: increasingly, custom software solutions and custom applications are the ones that truly solve complex problems in science and industry. Generic tools, while powerful, rarely adapt to the nuances of a particular domain, and DeepCormack is a crystal-clear example of how a bespoke approach can make all the difference.
This breakthrough also opens the door to new forms of collaboration between computational physics and the technology company. At Q2BSTUDIO, we understand that behind every scientific or business problem there is a flow of data that needs to be accurately modeled, processed, and visualized. Our expertise in artificial intelligence for companies allows us to accompany research centers and companies in the integration of AI agents that learn from synthetic or limited data, exactly as DeepCormack does. Whether it's optimizing rebuild routines in materials labs or deploying predictive models in industrial environments, the ability to build customer-specific solutions is our hallmark.
Beyond physics, DeepCormack's approach illustrates how artificial intelligence can rescue classic experimental techniques that seemed doomed to long measurement times. In fields such as cybersecurity, for example, similar principles apply: generating realistic synthetic data to train anomaly detectors when real attacks are rare or difficult to label. At Q2BSTUDIO, we offer cybersecurity and pentesting services that benefit from these same data augmentation techniques and supervised models, allowing vulnerabilities to be identified with an accuracy that previously required weeks of manual auditing.
Another fascinating aspect is the role of cloud infrastructure. Generating thousands of synthetic volumes and training deep networks such as DeepCormack's CNNs and UNet demand computational power that not all labs possess. This is where the AWS and Azure cloud services we deploy at Q2BSTUDIO come in: scalable platforms that allow researchers to launch massive workouts without investing in on-premises hardware, paying only for the actual compute time. In addition, integrating business intelligence services with Power BI can help visualize model evolution metrics during training or compare reconstructions of different materials interactively.
It's not just about speeding up science; it is about democratizing it. DeepCormack lowers the barrier to entry for studying Fermi surfaces in materials that were previously inaccessible due to the slow ACAR technique. Similarly, in the business world, automating processes using AI-powered automation solutions allows small and medium-sized businesses to access levels of efficiency that were previously only available to large corporations. At Q2BSTUDIO, we strongly believe in putting advanced technology at the service of every organization, whether it's a quantum physics lab or a logistics startup.
DeepCormack's potential is not limited to Fermi's surface. The methodology—which combines a classic physical model with a deep learning model trained on synthetic data—is transferable to other CT problems, such as characterizing defects in materials using X-rays or reconstructing medical images with low radiation doses. In fact, any domain where experimental data collection is costly or time-consuming can benefit from this paradigm. In Q2BSTUDIO, we are already exploring similar applications for our customers in sectors such as advanced manufacturing and energy, using AI agents that learn from simulations and are then deployed in real-world environments.
For this transfer to be effective, it is crucial to have a software ecosystem that integrates all stages: from data acquisition to deployment of the model in production. Here, the concept of tailor-made applications makes perfect sense. It is not enough to have a good algorithm; A software architecture is needed that handles data preprocessing, training generation, distributed cloud training, validation with experimental data, and visualization of results. At Q2BSTUDIO, we design complete solutions that cover each of these steps, ensuring that the final product is not only technically sound, but also usable by researchers who are not programming experts.
The future of material characterization and, by extension, of many scientific disciplines, lies in the synergy between artificial intelligence and classical experimental techniques. DeepCormack is a milestone on that path, but it won't be the last. As more groups adopt these types of strategies, we will see results times reduced dramatically, allowing for faster iterations in designing new materials with bespoke properties. At Q2BSTUDIO, we are prepared to accompany research centers and companies in this transition, offering not only technology, but also the necessary knowledge to adapt each solution to its specific context.
In short, DeepCormack represents a quantum leap in Fermi surface tomography, but its real lesson goes beyond that: it shows us that when you combine robust physical models with the flexibility of deep learning, and support all this with a scalable cloud infrastructure and custom software services, you can overcome limitations that seemed insurmountable. Whether at the frontier of scientific knowledge or at the core of a competitive business, the recipe is the same: understand the problem, customize the solution, and execute it with the most advanced tools. At Q2BSTUDIO, we do just that, day in and day out.
