iiinova — where your digital self keeps evolving

Xin@QuantumChem @xinquantum.bsky.social · 03/10/2026

Building catalytic surface using LightCone. Desktop version to be released with Direwolf (a quantum chemistry package). #compchemsky lightcone.quanta-bricks.com

051

Xin@QuantumChem @xinquantum.bsky.social · 24/09/2026

So called "molecule-molecule docking", it should be classified into conformer searching. Visualization: lightcone.quanta-bricks.com

000

Xin@QuantumChem @xinquantum.bsky.social · 03/05/2026

Yes, the data were released 7 years ago, so LLMs have likely been trained on them. I should use some private data to test a series of compounds.

000

Xin@QuantumChem @xinquantum.bsky.social · 03/05/2026

Data source: github.com/samplchallen...

github.com

000

Xin@QuantumChem @xinquantum.bsky.social · 03/05/2026

Testing three LLM models (Opus 4.7, Sonnet 4.6, Gemma 4) on LogP prediction. The pairwise ranking is the most important thing, 76% accuracy is acceptable. Accurate numerical prediction still relies on dedicated models, this test is more about how well LLMs understand chemistry. #compchemsky

200

Xin@QuantumChem @xinquantum.bsky.social · 23/04/2026

Catalytic surface editing is pretty easy on LightCone interface. Good for heterogeneous cat. system building. #compchemsky #compchem

010

Xin@QuantumChem @xinquantum.bsky.social · 13/02/2026

Two rescued cats. The first is named Isolated Belt (as I found it in an isolated belt bush). The second is Huahua (at least 9 yro). #compchem #biology #chemistry #compchemsky

140

Xin@QuantumChem @xinquantum.bsky.social · 20/01/2026

Extreme compression test on C60. Even under very aggressive deformation, xTB stays stable and gives a reasonable energy trend. (Without clear bond breaking). @myzzzz.bsky.social would like to share the detailed files.

030

Xin@QuantumChem @xinquantum.bsky.social · 28/11/2025

When it comes to real production level chemistry code, both Gemini 3 Pro and Claude Opus became clumsy. It took them around 40 minutes and multiple rounds of conversation just to finish the project. They’re clearly not great with RDKit or chemistry APIs. #CompChemsky #compchem #Chemistry

100

Xin@QuantumChem @xinquantum.bsky.social · 23/11/2025

In one sentence: Orb-v3 is accurate in describing F-Li-Be Molten Salt *without fine-tuning*. AIMD results came from: doi.org/10.1103/PRXE... They also compared with Neutron Diffraction. @myzzzz.bsky.social will benchmark the S(Q) with experimental data. #compchemsky #compchem

020

Xin@QuantumChem @xinquantum.bsky.social · 16/11/2025

I almost said protein folding was CompChem's holy grail, finally solved by AI. But I held back, just said: "It's AI predictions, double-check if they make sense."

010

Xin@QuantumChem @xinquantum.bsky.social · 16/11/2025

Wetlab is a different world. A wetlab biologist just told me she uses a visualization website where you input a protein sequence and click, the 3D structures show instantly. The visualization thing is called AlphaFold. #Biology #AlphaFold #compchemsky

140

Xin@QuantumChem @xinquantum.bsky.social · 09/11/2025

Centuries of human progress led us to uncover The Double Helix. In the end, we realized that all living things share the same codebook. #science #biology

010

Xin@QuantumChem @xinquantum.bsky.social · 02/11/2025

Yesterday I posted an incorrect comparison. I accidentally copied the TS3 structure into TS1 for UMA and Orb-V3, so the MLIP results for TS1 were wrong. Really sorry about that misleading! In the next post I will share IRC of DFT calculation and wB97M-V/def2-TZVPD results. #compchem #compchemsky

030

Xin@QuantumChem @xinquantum.bsky.social · 02/11/2025

Sorry this post is incorrect. I accidentally copied the TS3 structure into TS1. The corrected results should be:

100

Xin@QuantumChem @xinquantum.bsky.social · 02/11/2025

THIS POST is wrong: I wrongly copied the TS3 structure into TS1 in this post. The corrected results profile should be

000

Xin@QuantumChem @xinquantum.bsky.social · 02/11/2025

Oh, sorry, I made a mistake. I copied the TS3 structure into TS1. The corrected figure should be. As for wB97M-V/def2-TZVPD, its relative energy of TS1 is very close to M06-2X/def2-TZVPP (within <0.1 kcal/mol). I will share the notebook, IRC and data very soon.

121

Xin@QuantumChem @xinquantum.bsky.social · 02/11/2025

This is stricter than just comparing their own barrier height. Evaluating MLIP energies on the same DFT geometries, probe how similar the PES is. Answer the question: Given a decent DFT energy profile, can the MLIPs reproduce the relative energy along that path? It's crucial for the reactive MD.

100

Xin@QuantumChem @xinquantum.bsky.social · 02/11/2025

DFT was used to locate transition states and minima; UMA and Orb-v3 single-point energies were computed on the same structures with DFT.

000

Xin@QuantumChem @xinquantum.bsky.social · 02/11/2025

TS1 is located by DFT (black line). Red and green lines are from Orb-v3 and UMA.

000

Xin@QuantumChem @xinquantum.bsky.social · 02/11/2025

A failure case for deep-learning potential: the TS1 geometry is so weird that UMA and Orb-v3 underestimate the barrier by ~15 kcal/mol, likely because it far from the training set. Handling chemical reactions is extremely hard for general deep learning models. #compchem #compchemsky

542

Xin@QuantumChem @xinquantum.bsky.social · 22/10/2025

UMA (small model) is very close to DFT M06-2X. Maybe I can deploy nanoreactors using UMA model for coarse reaction space sampling. This is a reactive trajectory of a compound with 24 Carbon atoms. #CompChem #ChemChemSky

150

Xin@QuantumChem @xinquantum.bsky.social · 06/08/2025

playground.tracto.ai/playground/n... The newly released open-source model from OpenAI can be tried on TractoAI playground for *FREE*.

playground.tracto.ai

TractoAI | Cloud platform for AI and Data teams

Run AI and Data workloads in the cloud at scale

011

Xin@QuantumChem @xinquantum.bsky.social · 01/08/2025

This old theory is revitalized by deep learning. Building accurate kinetic energy functionals (< 0.5 kcal/mol) may be harder than climbing to the top of exchange-correlation Jacob's ladder.🪜

030

Xin@QuantumChem @xinquantum.bsky.social · 27/07/2025

Cunning ChatGPT bypassing 3D depth analysis, and calling the rdkit function to generate the coordinate of an Ibuprofen. Grok3 just failed this test and misrecognized it as glucose, as expected.

020

Xin@QuantumChem @xinquantum.bsky.social · 23/07/2025

Boltz2 can be run on a $430 RTX5060 Ti (16G) card, although it's much slower than ~$30,000 H100 (RTX 5060 Ti: 1min 47s, H100: 18s), but it doesn't matter, because the cold-start time is about 1 minute for both cards. #CompChemSky #CompChem #GPU

020

Xin@QuantumChem @xinquantum.bsky.social · 12/07/2025

This is a transition state structure (red is oxygen, blue is Nitrogen, Green is Nicle) of C-N splitting reaction. Create a deamon around the structure breaking the C-N bond.

010

Xin@QuantumChem @xinquantum.bsky.social · 12/07/2025

This is creepy, but lots of fun. Gemini Veo has a perception of spatial depth and can rotate the molecules. Inspired by www.youtube.com/watch?v=mDyt... #compchem #compchemsky

151

Xin@QuantumChem @xinquantum.bsky.social · 11/07/2025

It should be the warm start. AIMNet2's model downloading and loading are also fast. t0 = perf_counter() E=mol.get_total_energy() t1 = perf_counter()

000

Xin@QuantumChem @xinquantum.bsky.social · 09/07/2025

Correctness: AIMnet2 should be AIMNet2. Single-point energy calculation.

000

Xin@QuantumChem @xinquantum.bsky.social · 09/07/2025

Just run the following command in Colab. Works fine! !uv pip install --system git+https://github.com/aligfellow/vib_analysis.git

000

Xin@QuantumChem @xinquantum.bsky.social · 09/07/2025

I ran a benchmark on a protein with ~3450 atoms. AIMnet2 takes 4.6 seconds, and UMA(small) takes 7.9 seconds. CUDA 12.4 on a Tesla T4. Additionally, the AIMnet2 checkpoint file is only 11.1 MB. The testing notebook link is shared. #compchemsky #compchem colab.research.google.com/drive/1g1PiW...

colab.research.google.com

Google Colab

280

Xin@QuantumChem @xinquantum.bsky.social · 05/07/2025

This is a very interesting package. I will try this on colab notebook.

131