@thematterlab
The materials for tomorrow, today. We are the Matter Lab at the University of Toronto, led by Professor Alán Aspuru-Guzik. Our group works at the interface of theoretical chemistry with physics, computer science, and applied mathematics.
Colloquium: Professor Tim Newhouse of Yale University will discuss “Computationally Augmented Total Synthesis,” on Friday, March 13, 2026 at 10:00 am. Learn more: www.chemistry.utoronto.ca/events
From aging to fetal health and clean energy to pharmaceuticals, the Acceleration Consortium awards over $2 million to projects that accelerate scientific discovery in a wide range of fields. Read more: acceleration.utoronto.ca/news/from-ag...
Poster bearing the UofT department of Chemistry logo, a photo of a lecturer, and the same info, time/date and links as the text of this social post.
Physical Seminar Series: Gabriella Wang will discuss “Nonlinear Raman Response as Evidence of Photooxidation on the Surface of PbS Quantum Dots,” on Tuesday, March 10, 2026 at 11:00 am. Learn more: www.chemistry.utoronto.ca/events
The way to train the BioMetagenome and sequence embedding as BioNLP for large scale sequence inferences.
As tabular data and inferences.
You can buy one of these from a company in Singapore (exasynth.ai), light jazz not included: www.youtube.com/watch?v=gD1h...
Kudos to the authors: Raul Ortega Ochoa, @realmantilla.bsky.social, Juan Bernardo Pérez Sánchez, Mohsen Bagherimehrab, @an-aldossary.bsky.social, @tvegge.bsky.social, Tonio Buonassisi, and @aspuru.bsky.social.
[5/5]
Overall, the perspective aims to provide a more physically grounded framework for thinking about representation learning in chemistry.
[4/5]
The article highlights:
- Conceptual links between quantum tomography and modern molecular machine learning foundation models
- Informational completeness as a guiding principle for shaping latent structure
- Implications for dataset design, supervision strategies, and benchmarking
[3/5]
We introduce the Deep Tomography Hypothesis: as supervision becomes progressively more informative and approaches informational completeness, the space of admissible models reduces — encouraging representations that are more structured, constrained, and physically consistent.
[2/5]
Our perspective, Connecting the concepts of quantum state tomography and molecular representations for machine learning", is now published in Digital Discovery🎉
Here, we explore how ideas from quantum state tomography can inform representation learning in molecular ML
🔗 doi.org/10.1039/D5DD...
[1/5]
"A mobile robotic process chemist", just published in @digital-discovery.rsc.org pubs.rsc.org/en/content/a... Video shows 3 back-to-back autonomous reaction cycles (make product, analyse product (HPLC-MS), isolate solid product, clean reactor, check reactor is clean, loop) over 20 h. Reactor = 1 L
Kudos to the amazing team from UofT and NVIDIA for making this happen!
Luca Anthony Thiede, @an-aldossary.bsky.social, @andreasburger.bsky.social, Jorge Campos, Ning Wang, Alexander Zook, Melisa Alkan, Kouhei Nakaji, Taylor Patti, Jérôme Gonthier, Mohammad Ghazi Vakili, @aspuru.bsky.social
[7/7]
We believe MōLe lays the foundation for learning from molecular orbitals for a wide array of applications, such as higher-level CC methods, RDMs, quantum circuit parametrization, and others.
[6/7]
This is a step towards wavefunction-based ML, where the inputs (molecular orbitals) and outputs (excitation amplitudes) are both compatible with rigorous electronic structure theory.
[5/7]
🧪 More data-efficient than ∆-MP2 MLIPs, achieving chemical accuracy even with just 100 training molecules
⚛️ Accurate electron densities
[4/7]
⚡40–50% reduction in CCSD solver iterations and enabling convergence of hard-to-converge molecules when used as an initial guess
🚀 ~20× speedup vs CCSD in practice (with lower O(N⁵) scaling)
[3/7]
By predicting T1 and T2 amplitudes, MōLe (Molecular Orbital Learning) enables:
🎯 ~0.1 mHa energy error on QM7 (CCSD/def2-SVP)
🌐 Strong out-of-distribution generalization to molecules double the size of the training molecules & off-equilibrium geometries
[2/7]
Coupled cluster is the gold standard of quantum chemistry, but its steep scaling limits its routine use.
We introduce the MōLe model, the first equivariant neural architecture that directly predicts CC excitation amplitudes from Hartree-Fock molecular orbitals.
📃 arxiv.org/abs/2602.20232
[1/7]
Poster bearing the UofT department of Chemistry logo, a photo of a lecturer, and the same info, time/date and links as the text of this social post.
Colloquium: Professor Steve MacNeil, Wilfrid Laurier University "Why We (and Our Students) Should Not Be Content with Just Content: The Importance of Metacognition, Desirable Difficulties, and Productive Struggle in Higher Education", Friday February 27 at 10am.
To learn about the concurrently released El Agente Sólido, a new age(nt) for solid state simulations, check out this thread.
If you enjoyed this work, check out El Agente Gráfico, a structured execution graphs for scientific agents, released concurrently today.
bsky.app/profile/them...
Kudos to team, who made this possible: Sai Govind Hari Kumar, @yunhengzou.bsky.social, Andrew Wang, Jesús Valdés-Hernández, Tsz Wai Ko, Nathan Yue, Olivia Leng, Hanyong Xu, @ccrebolder.bsky.social, @aspuru.bsky.social, @variniabernales.bsky.social
[6/6]
We believe these results point to a promising direction for accelerating materials discovery by lowering the barriers to entry in computational materials science and making end-to-end simulation workflows more accessible to a broader range of researchers.
[5/6]
We evaluated the framework through extensive benchmarking exercises and case studies. Across seven benchmarking exercises, each iterated 10 times, El Agente Sólido achieved an average rubric score of 97.9% using rubrics designed by computational chemists.
[4/6]
It integrates density functional theory (DFT) with phonon calculations and machine-learning interatomic potentials, enabling simulations that are both efficient and physically consistent.
[3/6]
El Agente Sólido translates high-level scientific goals written in natural language into end-to-end computational pipelines, including structure generation, input-file construction, workflow execution, and post-processing analysis.
[2/6]
Next up, El Agente Sólido is a hierarchical multi-agent framework that automates solid-state quantum chemistry workflows using the Quantum ESPRESSO simulation package.
arxiv.org/abs/2602.17886
[1/6]
Kudos to the team who made this happen: Jiaru Bai, @an-aldossary.bsky.social, Thomas Swanick, Marcel Müller, Yeonghun Kang, Zijian Zhang, Jin Won Lee, Tsz Wai Ko, Mohammad Ghazi Vakili, @variniabernales.bsky.social, @aspuru.bsky.social
[7/7]
This is a step up from contemporary agentic frameworks, shifting the mindset from prompt engineering to context/harness engineering (with execution/knowledge graphs).
[6/7]
Not only that, but we also benchmark our system with 8 different LLMs. Across university-level quantum chemistry benchmarks, a single agent + structured execution engine can reduce the cost by ~96% compared to previous multi-agent systems, while achieving performance above 98%.
[5/7]
