Research in Machine Learning and Physics
Machine Learning String Theory
String Theory has revealed to be the most promising candidate for a unifying Quantum Gravity theory. Allegedly, within String Theory are encoded the laws that regulate our Universe. However, String Theory is a complicated, entangled beast, and it is not so straightforward to tell whether String Theory truly explains the Physics of our Universe. Indeed, firstly, String Theory predicts a gigantic amount of models that could explain our Universe, and it is hard to pick which is the correct one that is fully consistent with our observations. Secondly, String Theory models are based on sophisticated geometrical constructions, endowed with a huge number of parameters to be tuned.
In recent years, it has become clear that the huge complexity of String Theory can be better addressed statistically, rather than systematically, by employing Machine Learning techniques. Currently, my research mostly revolves around this topic, to which I am contributing following two directions:
The theoretical approach: what we can learn in Quantum Gravity
Although Neural Network-based Learning can be crucial for learning several predictions of String Theory, it is not clear what can we actually learn or to what extent we can learn something. For instance, it is crucial that compelling problems such as does String Theory allow for Inflation? or does String Theory include the Standard Model? are learnable. On theoretical grounds, I explore whether key problems such as these can be learned. To this end, I utilize concepts and theorems from Statistical Learning Theory, and see how these connect with the geometrical structures of the String Theory-originated effective theories.
As a first step in this direction, in this work
- Neural Network Learning and Quantum Gravity
Stefano Lanza
Published in: JHEP, vol. 07, p. 105, 2024
PDF, arXiv, INSPIRE-HEP
I have illustrated that the geometrical structures that characterize Quantum Gravity imply the statistical learnability of several problems that one can therein formulate.
The practical approach: how we can learn features of Quantum Gravity
On practictal grounds, I am applying Machine Learning techniques to address compelling problems in String Phenomenology. This approach serves, on the one hand, to unearth new, undiscovered properties and patterns of String Theory effective theories via the power of Machine and Deep Learning techniques; on the other hand, such a practical approach is useful to corroborate and substatiate the findings of the above theoretical viewpoint.
Along this direction, in the following work
- Machine learning the breakdown of tame effective theories
Stefano Lanza
Published in: Eur. Phys. J. C, vol. 84, no. 6, p. 631, 2024
PDF, arXiv, INSPIRE-HEP
I have employed supervised classification algorithms - the k-nearest neighbor and the Linear Support Vector Machine algorithms - in order to predict the kinds of states that break-down stringy effective field theories.