Max Welling

Max Welling is a machine-learning researcher and co-founder of CuspAI. His joint contributions include variational autoencoders, group-equivariant convolutional networks and scalable graph convolutions. In materials discovery, he advocates combining generative models, simulations and physical experiments into tools that accelerate chemists’ work while keeping domain experts involved.[3][4][5][6][7][1]

Welling and Chad Edwards founded CuspAI in 2024 to build AI systems for materials discovery. In an interview published February 25, 2026, Welling explained that scientific curiosity had increasingly been joined by a desire for practical impact, particularly because of his concern about climate change. He saw materials science as combining interesting scientific problems with opportunities to build useful technology.[7][1]

Welling described a platform architecture he had sketched at the outset: a generative model proposes candidates, then a multi-scale, multi-fidelity digital twin filters them through progressively more expensive evaluations. Cheap checks eliminate unsuitable candidates first; a smaller set reaches experiments, whose results provide feedback. He said multi-scale modeling, laboratory automation and agents had become additions to the design, with agents searching chemical literature and orchestrating computations and experiments at varying stages of maturity.[1]

Welling emphasized that constructing the platform was harder than outlining its architecture, and credited collaborators responsible for its scientific and platform/MLOps components. CuspAI identifies Felix Hanke as VP Scientific Engineering and Alessandro De Maria as VP Platform Engineering.[1][7]

His development approach starts with chemists assembling workflows manually from modular tools, then automating portions once the required steps are understood. As an illustration, he described a porous material that collapses when shaken: the workflow needs an additional stability test before an agent can reliably use it. Automation can then grow from smaller tasks into larger sequences, using agents or Bayesian optimization.[1]

Welling calls physical experimentation a “physics processing unit”—“nature doing computations for you”—that could work alongside digital computation. His longer-term vision is a materials search engine: a chemist specifies desired properties, receives candidates and judges which are useful. He does not expect domain experts to leave the loop soon. Moving to a different materials problem can require different experimental equipment, retrained models and renewed expert guidance.[1]

Welling also described requiring an industrial partner before investing in a materials direction, because bringing a material into use demands domain expertise. His strategy combines an undisclosed ambitious materials goal with more modest paid projects, such as training a force field or analyzing a specific problem. He expressed a preference for deep, long-term partnerships and hoped for a breakthrough material before complete automation.[1]

This materials work follows an earlier career in theoretical physics, probabilistic learning and neural-network architecture. Those earlier contributions provide context for the generative models and symmetry-based methods Welling discusses in his materials work.[8][3][5][1]

Welling earned a PhD in theoretical high-energy physics at Utrecht University under Gerard ’t Hooft, then transitioned to AI during postdoctoral work at Caltech and later with Geoffrey Hinton, according to his May 2025 NYU seminar biography. As of that seminar, he held a professorship and research chair at the University of Amsterdam and served as a Merkin Distinguished Visiting Professor at Caltech.[8]

Welling and Diederik P. Kingma co-authored Auto-Encoding Variational Bayes, first submitted on December 20, 2013. Their method addresses learning and inference with continuous latent variables and intractable posterior distributions. Under differentiability conditions, reparameterizing a variational lower bound produces an estimator that standard stochastic-gradient methods can optimize. For independent, identically distributed data with continuous latent variables per datapoint, a learned recognition model efficiently approximates posterior inference.[3]

The paper won the inaugural ICLR Test of Time award in 2024. The organizers credited it with integrating deep learning and scalable probabilistic inference, giving rise to the variational autoencoder. They also acknowledged concurrent work by Rezende and colleagues, published at ICML 2014, that proposed a similar idea.[4]

In 2013, Welling co-founded Scyfer with Taco Cohen, Tijmen Blankevoort and Jorgen Sandig. The Amsterdam spin-off applied machine learning to industrial problems, including a Tata Steel inspection algorithm for detecting defects in steel plates. The university announced Qualcomm Technologies’ acquisition in August 2017. Welling went on to lead AI research at Qualcomm for four years before joining Microsoft Research as Vice President and Distinguished Scientist, according to his May 2025 NYU seminar biography.[9][8]

Cohen and Welling’s Group Equivariant Convolutional Networks, published at ICML 2016, exploit symmetries to reduce sample complexity. Their group-convolution layers share weights more extensively than ordinary convolution layers, increasing expressive capacity without increasing parameter count. The paper describes implementations for discrete groups generated by translations, reflections and rotations. Looking back in the 2026 interview, Welling credited Cohen as the main driver of their symmetry work, which subsequently extended toward gauge symmetries.[5][1]

Welling illustrates the benefit of built-in symmetry with a rotated bottle: training in one orientation can carry over to others, reducing the data needed. He also stresses the trade-off. Architectural constraints can complicate optimization, and data augmentation can sometimes work better. An incorrect inductive bias can impose a performance ceiling; he argues that architectures must also be able to scale.[1]

With Thomas N. Kipf, Welling co-authored Semi-Supervised Classification with Graph Convolutional Networks, first submitted in September 2016 and published at ICLR 2017. Their architecture uses a localized first-order approximation of spectral graph convolutions. It learns representations combining local graph structure with node features and scales linearly in the number of graph edges.[6]

In the February 2026 interview, Welling described completing a manuscript called Generative AI and Stochastic Thermodynamics and sending it to a publisher. He said it developed mathematical connections between generative models and non-equilibrium physics, including diffusion models, free energy and sampling. He viewed those connections as opportunities both to improve learning algorithms using physics and to help scientists through machine-learning models.[1]