.jpg)
The discovery of new materials has historically been a costly and slow process, where the experimental or computational evaluation of properties (such as thermal conductivity or elastic modulus) consumes most of the resources. In recent years, artificial intelligence has revolutionized this field through generative models that propose promising crystal structures, but the bottleneck remains validation with high-cost simulations (e.g., DFT). An emerging strategy consists of interposing a probabilistic surrogate —a fast machine learning model— that acts as an intelligent filter between the generator and the expensive oracle. This Bayesian design of materials employs a Gaussian process as a gate that scores and ranks candidates, allowing the oracle to focus only on the most promising ones. Results show that, with a limited evaluation budget, using a surrogate based on pre-trained embeddings (such as ORB) can match or even exceed the performance of direct fine-tuning of the generative model, reducing oracle calls by a factor of five with minimal loss of precision. This approach not only accelerates discovery but also opens the door to tailored applications in sectors such as energy, electronics, or catalysis.
The key to success lies in how the surrogate learns to represent the generated structures. Instead of relying on manual descriptors, embeddings extracted from neural networks pre-trained on large crystal databases are used, capturing complex relationships between composition, symmetry, and properties. Combined with a Gaussian process, the model provides not only a point prediction but also an associated uncertainty, allowing selection criteria such as ranking to be applied, which has proven to be much more effective than random selection. This type of workflow is perfectly transferable to the business environment, where companies need to make quick decisions based on limited data. Implementing such a system requires custom software that integrates generative models, probabilistic surrogates, and oracles, all orchestrated to maximize efficiency without sacrificing quality.
At Q2BSTUDIO, as a software development and technology company, we understand that the integration of artificial intelligence into research and development processes must be robust and flexible. Our services include creating platforms that combine AI agents for candidate generation, machine learning-based surrogates, and scalable evaluation systems. Additionally, we leverage AI for businesses to design pipelines that adapt to specific domains, whether in materials, pharmaceuticals, or logistics. The cloud plays a fundamental role in this context: AWS and Azure cloud services allow deploying massive simulations and storing high-dimensional embeddings without worrying about local infrastructure, while cybersecurity ensures the protection of proprietary data. All of this can be complemented with business intelligence services, such as Power BI, to visualize in real time the progress of discovery campaigns and surrogate metrics.
One of the differentiating advantages of the Bayesian approach is its ability to operate with very limited evaluation budgets. This is especially relevant in industrial environments where each simulation costs time and money. By using a surrogate that ranks the generator's proposals, computational effort can be concentrated on the few structures that truly deserve validation. For example, in the search for new battery electrolytes, a generative model can produce thousands of candidates, but only a few dozen will reach the oracle, and of those, a fraction will be tested experimentally. This type of optimization, which combines machine learning with experimental design, is a perfect application of the concept of AI agents working in a chain: the generator agent proposes, the surrogate agent filters, and the oracle confirms.
The practical implementation of these systems requires a multidisciplinary team that masters both materials science and software engineering. At Q2BSTUDIO, we offer custom application development services that allow personalizing each step of the workflow: from choosing the most suitable embedding to integrating with external oracles (DFT, MD simulations, etc.). Our experience with AWS and Azure cloud services ensures the solution is scalable and maintainable, while the business intelligence cyber-services layer provides visibility into the process. If your organization seeks to accelerate materials discovery or any other domain where evaluation is costly, a probabilistic surrogate approach with pre-trained embeddings may be the answer, and we can help you bring it to life.
