At the same time, a top-quality datasets are essential to make physics-AI models truly predictive. That’s why so much of our focus is on creating and curating diverse datasets across chemistry, materials, and biochemistry.

I discussed why physics-based AI models remain indispensable in chemistry and materials discovery—purely data-driven AI will never match the reliability of physics-AI.

I’m grateful to Bakar Bio Labs for the opportunity to share what we’re building at Azulene Labs and to be part of such a vibrant community.

#BakarLabs #UCBerkeley #DeepTech #Chemistry #MaterialsScience #Biochem #Biotech #PhysicsAI #AIforScience #MaterialsDiscovery #MolecularDiscovery

Azulene Labs is #hiring ✨ www.linkedin.com/posts/nicola...

#compchem #aiforscience #computationalphysics #physics #molecularmodeling

Models like Boltz-2 don’t replace expensive physics-based calculations—they allow you to do fewer of them. The practical workflows will be to use Boltz-2 (or similar) for an initial screening, and proper FEP for a "second pass" step.

#ComputationalChemistry
#DrugDiscovery
#MachineLearning
#Boltz2

One note: in these benchmarks, FEP+ uses (I think) an OPLS-based force field, which itself can/will be improved upon with more modern force fields. So as force fields improve, the gap between Boltz-type models and physics-based methods will likely get even larger.

On their test set, Boltz-2 showed a Pearson R of 0.63, compared to 0.72 for FEP—extremely impressive but still a large gap. The speed of Boltz-2 comes at the cost of accuracy, and if the baseline FEP model isn’t accurate enough to begin with, that tradeoff isn’t always justified.

But it’s important to recognize, as the authors themselves make clear, that Boltz-2’s accuracy is still far behind expensive physics-based FEP (Free Energy Perturbation) methods.

Boltz-2 made waves in the chemical modeling world last week. It’s a big leap forward—very fast binding energy predictions with decent accuracy. It’s probably capturing some entropic effects implicitly, which is genuinely exciting.

Excited to announce that Azulene Labs is teaming up with the Young Professionals in Energy (SF chapter) and Dunia to host a special Molecules & Materials Meetup—part of Deep Tech Week. 🚀

Space is limited—please register here: lu.ma/2dpllaxe

#DeepTechWeek #Chemistry #MaterialsScience #BioTech

Bottom line: VQE is here to stay, and will be at the heart of tomorrow’s most important quantum algorithms. 🚀

Read full post here: www.nicolassawaya.com/post/vqe-isn...

#QuantumComputing #QuantumAlgorithms #DeepTech

All the work being done today on variational algorithms, optimization, and ansatz design is laying the foundation for the algorithms we’ll run on error-corrected quantum computers. In fact, removing noise constraints will make these methods even more powerful and precise.

The reality: VQE will be an *indispensable* *subroutine* in future quantum algorithms. Whether you’re doing QPE, Krylov subspace methods, or imaginary time evolution, you need a good starting state. VQE delivers that, making it a key first subroutine—not just a noisy hardware workaround.

Quick post on quantum algorithms here ⚛️

There’s a common misconception: VQE (Variational Quantum Eigensolver) is “just a NISQ-era algorithm” that will be obsolete once we have fault-tolerant quantum computers. I hear this sentiment from far too many people. It’s simply not true.

Earlier this year we activated noether1 through noether8. These eight nodes are named after Emmy Noether, the revolutionary mathematician.

This expansion accelerates our creation of the world's most predictive AI models in chemicals & materials 🧠

#HPC, #AI, #Science, #MolecularDesign