It is by now widely-accepted that artificial intelligence (AI) has the potential to change scientific research as we know it. In fact, it is believed that a kind of a more specialized version of AI, one that mimics the scientific method, might be a goal easier to achieve than that of a general-purpose AI. This belief is based on the fact that, unlike other aspects of human life, scientific procedures are based on precise well defined rules and logic. A particular type of artificial neural networks, known as generative models, carries the potential to achieve this goal.

Scientific discoveries typically result from the work of human researchers by following the scientific method. It all starts with a question, the study of which typically involves a long chain of trial-and-error steps. In each of these steps, different hypotheses are created and tested. In many cases, this can be a painstakingly long process and scientists spend months or even years working on a single question without finding a satisfying answer. One of the problems is that the space of possible answers is simply too large to be systematically explored by human scientists. In fields such as drug discovery, for instance, it’s believed that there are about 1063 different molecules that could be used as the basis of medical drugs [1].

Scanning all possibilities via a trial-and-error approach does not seem to be a reasonable option. Computers on the other hand have the ability to perform evaluations very fast and have been proven to be very useful in the span of the last 80 years, helping scientists automate substantial parts of their work. In traditional scientific research, along with experiments and observations, computer simulations help scientists evaluate the validity and potential of different hypotheses assumptions. However, the conceptualization of the hypotheses themselves has always been exclusively a human job. It is believed that human-like traits, such as creativity and inspiration, are prerequisites for undertaking this step.

A category of AI algorithms, the so-called generative models (GMs) [2], is designed to perform a very particular type of task: to generate artificial but realistic data based on large sets of observations. For example, a GM can be trained on signals outputted by a detector and learn to produce new artificial events. If enough data has been used for training and the level of complexity of the model matches that of the underlying physical mechanism that produces the measured signals, these artificial events will preserve all statistical and physical characteristics of the original ones. This is achieved by trying to approximate the underlying true distribution, that summarizes the characteristics of the physical mechanism, hidden behind the real events. At a first glance, this might be mistakenly thought of as simply a different kind of computer simulation. However, the differences are striking. Simulations are based on theory and assumptions (hypotheses) about the world, while in the case of GMs, no such assumptions are required. The fact is that GMs are based solely on observations and are able to uncover an approximate distribution that fully captures the essence of a physical mechanism. This distribution can be seen as an artificially-created hypothesis. So, in a way, trained GMs are in fact hypothesis-generating machines.
The most popular and highly successful deep neural-network architectures for training GMs are the so-called generative adversarial networks (GANs) which are based on a zero-sum game between two machine players, a generator, who generates new artificial events, and a discriminator, who evaluates the quality of the produced samples. However, alternative approaches for creating GMs do exist. Network architectures, such as variational autoencoders or normalizing flows are some examples. GM methods have been proposed in many scientific fields for the creation of artificial scientific hypotheses. In theoretical physics, string theorists have proposed the use of GMs for making approximate predictions in the string theory landscape [3]. In experimental particle physics, researchers use GMs to create artificial particle collision events [4]. In astrophysics, GMs are used to explore the evolution of galaxies [5] and better understand dark matter [6]. In drug discovery, GANs are already being used to generate novel molecules and build a virtual molecule library [7]. And in the field of clinical medicine, researchers have experimented with GMs to be used in so-called in-silico clinical trials [8], which, in many cases, allow the dramatic acceleration in the development of new drugs and medical devices and at the same time significantly cut R&D costs.

In the future, GMs will most probably not be limited to the creation of plausible scientific hypotheses. They are expected to help with even the most fundamental step towards a scientific discovery: proposing the scientific question itself.

[1] the art and practice of structure-based drug design: a molecular modeling perspective
R. S. Bohacek, C. McMartin & W. C. Guida
Medicinal research reviews 16, 3 (1996)

[2] deep learning
I. Goodfellow, Y. Bengio & A. Courville
MIT press 2016

[3] statistical predictions in string theory and deep generative models
J. Halverson & C. Long
Fortschritte der Physik 68, 2000005 (2020)

[4]event generation with normalizing flows
C. Gao, S. Höche, J. Isaacson, C. Krause & H. Schulz
Physical Review D 101, 076002 (2020)

[5] exploring galaxy evolution with generative models
K. Schawinski, M. D. Turp & C. Zhang
Astronomy & Astrophysics 616, L16. (2018)

[6] fast cosmic web simulations with generative adversarial networks
A. C. Rodriguez, T. Kacprzak, A. Lucchi, A. Amara, R. Sgier, J. Fluri, T. Hofmann & A. Réfrégier
Computational Astrophysics and Cosmology 5, 1 (2018)

[7] in generative adversarial learning: architectures and applications
Z. Zhang, F. Li, J. Guan, Z. Kong, L. Shi & S. Zhou
Springer, Cham.: GANs for molecule generation in drug design and discovery 233 (2022)

[8] artificial intelligence for in silico clinical trials: a review
Z. Wang, C. Gao, L. M. Glass, & J. Sun
arXiv preprint, arXiv:2209.09023 (2022)