The rapid progress of Artificial Intelligence (AI) is transforming nearly every facet of scientific research, with Quantum chemistry and electronic structure theory not being exceptions. Density Functional Theory (DFT) [1–5], the workhorse of these disciplines, is also undergoing substantial evolution under the influence of AI, leading to significant impacts on molecular and material simulations at various scales [6–13].

Specifically, AI’s involvement in DFT serves two main purposes: Firstly, it is used to improve [1–9] and accelerate [2,10–13] DFT approximations, thereby striving to resolve or at least alleviate [14] the problems associated with functionals designed by humans. Secondly, AI is being employed to create surrogate models that reproduce DFT results  [15–22,22–24].  In between these two strategies, there are also “Δ-machine learning” approaches, where (machine learning) ML is used to refine properties derived from DFT and bring them close to wavefunction accuracy [25–27]. Such models facilitate simulations on larger lengths and time scales [18,22]. The fusion of DFT and AI is also revolutionizing chemical research by accelerating high-throughput screenings by several orders of magnitude [20,28,29].

In this workshop, we will concentrate on the latest advancements at the intersection of AI and DFT, with a specific emphasis on two main areas: AI-powered improvements of DFT and the development of AI surrogate models aimed at reproducing DFT results. Our goal is to foster knowledge exchange between experts working on these interrelated areas and attempt to answer and better understand the implications of the following open challenges in the two fields:

1.         When and how will AI-learned functionals attain the broad applicability that their human-designed counterparts already have?

2.         Can the strong correlation problem in DFT [30–34] be mitigated or even solved by AI?

3.         How is the lack of variety and reliability of benchmark data for extended systems, and for transition metal molecules [35–37], hampering the evolution of AI-based DFT methods? What solutions can be applied to rectify this?

4.         Can AI enable further progress in handling dispersion interactions, beyond the standard heuristic corrections such as those introduced by Grimme [38,39]?

5.         What strategies can be used to further advance ML-based DFT approximations beyond their current state-of-the-art? For example, what is next after DM21 [8]? 

6.         Can the use of novel DFT features advance ML-DFT [32,33]?

7.         What are the key challenges in creating Machine Learning models for real-world applications (e.g., catalysts discoveries) that are based on the DFT data [28,40,41]? For this question, we aim to aim to distinguish and address the challenges that arise from DFT itself, and those that are related to the ML models. Moreover, we aim to distinguish and address the challenges that arise from DFT itself, and those that are related to the ML models.

8.  In what way can ML models be informed by the "zoo" of DFT functionals rather than inheriting the bias of a single one?