Mark Wronkiewicz

💡 ByteDance's Latest Work on Neural Network-Based Molecular Force Fields 💡 I recently read ByteDance’s work on developing a new force field, and it’s pretty good! 👍 ByteFF is a molecular mechanics force field trained on a massive dataset: 2.4 million geometries and 3.2 million torsion profiles. It leverages GNN to predict bonded and non-bonded parameters, showing strong performance in predicting molecular geometry, torsion energy, and conformational energy. This model has great potential for drug discovery and molecular simulation. I do have some concerns about using torsion scan + hessian methods to train non-bonded parameters in this framework. Non-bonded interactions often involve long-range effects, which are challenging for GNNs to capture accurately. That said, the overall results presented in this work are still quite promising. It's exciting to see tech giants like ByteDance stepping into AI for Science, and their vast resources enable them to support DFT calculations on this massive scale. (🤞 Hopefully, they will release the dataset for open use, LOL!) 👉 You can find the paper here: https://lnkd.in/gjmGV4fd 👉 And check out their open-source framework on GitHub, which is well-engineered: https://lnkd.in/gkxcTi6t #AIforScience #DrugDiscovery #MolecularModeling #ForceField #ByteDance

49

1 Comment

Heliophysics intersects so many things. Here is a key insight from decades old data gained because this team put their knowledge of planetary systems in the context of heliophysics. Also, love that the Voyager data continues to deliver! https://lnkd.in/eiCnnceC

61

2 Comments

🌟 Revolutionizing Structural Elucidation with AI! 🌟 For decades, computer-aided structural elucidation (CASE) for organic compounds has been bound by complex, rule-based systems, limiting their efficiency with intricate molecules. I’m thrilled to share our latest work on advancing this field by introducing a new transformer-based, generative AI model as a fresh approach to spectroscopic-based structural elucidation. Our proof-of-concept model harnesses the power of an encoder-decoder architecture and self-attention mechanisms, akin to those in large language models, to directly generate chemical structures from spectroscopic data—eliminating the need for traditional CASE logic. ⚡ Fast and Accurate: Trained on a dataset of ~102k IR, UV, and ¹H NMR spectra, our model can predict molecular structures with up to 29 atoms in just a few seconds on a standard CPU, achieving a top-15 accuracy of 83%. 💡 Why This Matters: By streamlining the structural elucidation process, this model paves the way for rapid advancements in molecular analysis. Its capability to quickly adapt to new data suggests a transformative potential in accelerating scientific discovery in chemistry. Excited about the possibilities that transformer-based generative AI holds for scientific problem-solving! Read more in our publication here: https://lnkd.in/e7sMQ_xv #ArtificialIntelligence #MachineLearning #Chemistry #StructuralElucidation #Innovation #ScientificResearch #GenerativeAI #Transformers

20

1 Comment

🌟 Ever wondered how we can decode fast limb movements using brain-computer interfaces (BCIs) with slow data acquisition, like in optical imaging? This challenge has long puzzled researchers, but Seungbin Park and her team, led by Dr. Maria Dadarlat at Purdue University, have come up with a solution! 🎙️ In this episode of the BCI Award Neurocareers podcast series, Milena Korostenskaja, PhD interviews Seungbin Park to discuss the problem of low sampling rates in decoding limb movements. While traditional methods struggle to capture fast limb trajectories, Seungbin and her colleagues have used a deep learning approach that overcomes this barrier! 💻 By employing a recurrent encoder-decoder network, they successfully decoded rapid limb movements of a running mouse using slow two-photon calcium imaging—a breakthrough in minimally-invasive, optical imaging-based BCI for multi-limb control. 🧠💡 🏆 Join us as Seungbin shares how their innovative approach earned them a spot as one of the 12 finalists for the BCI Award 2023 sponsored by g.tec medical engineering GmbH - NEVER STOP RECORDING. She’ll also offer insights and advice on how to make a successful BCI Award submission! Also, get resources and advice to help you start applying deep learning in your neurotech work (make sure to look into the podcast notes for useful links)! ✨ Tune in to hear Seungbin Park’s incredible story of pushing the boundaries in neurotech! https://lnkd.in/eFfpcuK5 🚨 Exciting News! 🚨 The BCI Award Foundation has just announced the 12 nominees for the prestigious Annual BCI Award 2024! Curious? Find out who was nominated here: https://lnkd.in/ds9BxUY 🏆 Save the Date: Join us for the BCI Award ceremony on October 8, 2024, where all BCI Award nominees will present their projects, and the BCI Award winners will be announced! https://lnkd.in/eU43rdhF #neurotech #neuroscience #braincomputerinterface #bci #opticalimaging #bciaward #stem #award #deeplearning #machinelearning #artificialintelligence #computerscience #engineering #technology #engineeringtechnology #brain #brainscience #neuroapproaches #neurocareers #podcast #neurocareerspodcast #careeradvice #research #careersuccess #careerdevelopment #brainresearch #neuralnetworks

24

1 Comment

I'm thrilled to share my latest personal passion project: FindingEdoc (ask me about my love for Stephen King's works)—a chatbot designed to help developers explore and understand codebases more efficiently. By leveraging generative AI, knowledge graphs, and Retrieval-Augmented Generation (RAG), this tool streamlines the coding experience when working with new or custom projects. Check it out here: https://lnkd.in/g6PgfVbc Throughout this journey, I’ve learned a ton: - Using LLMs and Neo4j to augment structured data creation. - Harnessing LangChain to retrieve meaningful context. - Simplifying interaction with Docker and Gradio. - While human coding jobs are safe from LLMs, I found joy (and pain) that the real bugs were mine all along! I’m excited to keep refining this solution, and I'd love to connect with anyone interested in knowledge graphs, human-AI collaboration, and RAG workflows. #AI #KnowledgeGraphs #LangChain #Neo4j #GenerativeAI #RAG #HumanAIteaming

16

For MRI folks: we just rolled out a new release to torchkbnufft, first in a couple years. The changes are for working with newer package versions. Things now work on numpy 2.0, and a few deprecations are fixed. Other than that, it's the same as before :). Get it with `pip install torchkbnufft` https://lnkd.in/ezyXrj-q

54

Software Engineering and High Performance Parallel Computing: As I’ve recently discussed, modern hardware is built with multiple cores per processor, and to be efficient, modern software should utilize those cores. For example, your laptop might be running 100 processes right now as you are reading this post. Most of those processes are lightweight so they can idle in the background as your CPU task switches between active and inactive ones. 100 processes on 4-12 cores is no problem when those processes spend a large amount of time idle, but in high performance computing, our processes are not idle. Running 2 processes on the same core causes contention and offers no speedup. When designing scientific codes, it is more appropriate to maintain a “1 process per core, or 1 data per core” mental model. We are trying to maximize the utilization of cores as we spray data across them. What happens when the size of our data is too large to process on a single machine? Well, that is where these massive parallel processors (MPPs / HPCs). In this model, we are actually sharing data in memory between machines. But how do we do that? Specifications like Open Multi Processing (open MP) allow you to parallelize tasks over a multicore processor when they have shared memory between them to simplify parallelism, but they work with shared memory on a single processor. Message Passing Interface (MPI) is a specification that allows you to parallelize code over a distributed system, such as networked computers in a supercomputer which each contain their own memory and cores. These are specifications that have language implementations in common scientific computing languages like C, C++, and Fortran. At the most basic level, MPI allows for distributed processing via message passing as the name implies, and a basic MPI implementation must be able to do the following 4 things at minimum: 1. Start processes 2. Send messages 3. Receive messages 4. Synchronize data / processes With these 4 capabilities, we can construct a distributed MPI program. MPI as an API offers over 400 functions, but with 10-15 we can implement the majority of algorithms we’d wish to run across large systems. For a basic introduction to writing some MPI code, check out this resource: https://lnkd.in/eA-eqcBm. If you like my content, feel free to follow or connect! #softwareengineering #parallelcomputing

131

9 Comments

Please see our new paper: Deep Adaptive Phase Learning (DAPL). We created a way to *coherently* autofocus systemic phase error in synthetic aperture sonar images using deep learning. Our method removes the need for hand-crafted windowing and convergence heuristics like those needed in methods like PGA, along with the need for specific image content. Likely, this can be applied to SAR imagery as well as our method works on the single-look-complex (SLC) image! https://lnkd.in/dckFqAbA

42

1 Comment

If you are just getting started with visually-guided navigation, I am sure this article will give you a good overview of the process in a systematic way with some practical considerations. https://lnkd.in/d4eDuBTN

11

Detecting Primordial Black Holes: New Methods ✦ https://lnkd.in/grzGcrXZ Primordial black holes (PBHs), hypothesized to form after the Big Bang, intrigue scientists as dark matter candidates and sources of gravitational waves. Current research explores using neutron stars, planets, and metal plates to detect PBHs by identifying microchannels they might leave, potentially confirming their existence. https://lnkd.in/grzGcrXZ #astronomy #space #physics

1

Here is a recent commentary I wrote about a recent science discovery by Li et al., enabled by the NASA Magnetosphere Multiscale (MMS) mission and its implications for solar wind, heliospheric and magnetospheric physics. To understand the compelling physics, solar wind structure and quantify the role of various cross-scale processes on particle transport and acceleration, we need a dedicated Solar Wind Multi-Scale mission in near future! https://lnkd.in/ezNkYQwU

35

Read more about the research of UC Merced professor and CITRIS PI Shijia Pan in this featured profile:

4

Mechanical engineering researchers from UIUC and Santa Clara recently published a research study in Advanced Science. Their study, “Curvature-Assisted Vesicle Explosion Under Light-Induced Asymmetric Oxidation," combines experiments and modeling to explore vesicle dynamics in nonequilibrium conditions. This work offers insights into biomembrane stability under oxidative assault and informs strategies for controlling release dynamics in vesicle-based drug delivery systems. Learn more in the press release here: https://lnkd.in/eGy9mp7V #SCUEngineering #MechanicalEngineering

17