The new language model for science . Upon few quick tries, it seems to generate professional text in the areas I am familiar with. And 7 years ago we were *joking* about ML writing papers!
HUGE CONGRATULATIONS to our 2022 Foresight Institute Feynman Prize Winners! Sergei V. Kalinin
@Sergei_Imaging
for Experimental Work, James R. Chelikowsky for Theory, and Dr. Emanuele Pennocchio
@EPenocchio
Distinguished Student Award 🎉
@tunguz
As proposed by
@MaximZiatdinov
, there is no contradiction between code and paper. They can be combined in a single format, with text as markdown and code as code. Plus all the refs, etc. Then this document can be cited as paper, and code can be shared
Rama Vasudevan
@ramav_matsci
-> APS Fellow, "For pioneering and visionary development of open-sourced physics-based machine learning methods in atomic-scale and mesoscopic imaging, and their application in physics". Congratulations, Rama!
DFT gives some (often good) approximation of reality. ML on DFT allows to interpolate, but it's approximation of approximation. The challenge is for ML to update ML/DFT models based on experimental data.
Dear colleagues - if (a student) or (have a student) you would like to spend a year at ORNL to work on science driven machine learning based automated experiment in electron or scanning probe microscopy - let me know (check out Office of Science Graduate Student Research (SCGSR).
New experience in scientific life - have paper rejected from a journal ( transfer offer), submit it to a different one and get very positive reviews, and get (out of the blue) acceptance (!) from the first journal.
First time I get review that says "This is a well-conceived, timely, and visionary article whose publication will be a central part of the revolution of elevating automation from a local optimizer to a global accelerator of scientific discovery”
SVK citations -> 45k. Always interesting to explore areas between disciplines - nanoscale electromechanics, machine learning in imaging and materials design, and direct atomic fabrication. And for all three of them the next decade looks like becoming mainstream!
As a professional recognition - I am honored to be invited to give the Manuel Cardona lecture of the The Institut Català de Nanociència i Nanotecnologia, ICN2.
SVK Career update (belated): ORNL -> Amazon (principal scientist, sabbatical)/University of Tennessee Knoxville (professor). My new email is sergei2
@utk
.edu
Looks like there starting to be a very strong resonance to recent Nature papers by Google Brain team and A-Lab.
One immediate thought:
- we as a community really invested a lot in theory
- and not really enough in inverse problems - analysis of scattering and spectral data
For ML, the possible next step beyond status data sets and SOTA chasing may be deployment on (almost) closed experimental systems exploring or optimizing materials with relatively simple physics including automated microscopes or automated synthesis, as
Atomic-scale e-beam sculptor patent granted!
Took a while after it was filed - covers atomic motion, 3D crystallization, and feedback based on learned cause and effect relationships between beam and atomic changes.
@MaximZiatdinov
Why? If followed, this principle will really open doors to very unfortunate outcomes (suppress collaboration, dirty politics, you name it)
We argue that the unique opportunity given by electron and scanning probe microscopies is to learn generative physical models (much like astronomers create models of universe based on observations).
@ramav_matsci
@MaximZiatdinov
Enter the hypothesis learning - aka active learning co-navigating theory and experiment via hybrid of reinforcement learning and structured Gaussian Processes
@MaximZiatdinov
And this is the condition they Steve Pennycook, the best mentor I had when at ORNL, made me sign as part of a deal to collaborate on STEM on functional materials. Hopefully 18 years later, ChatGPT will do it for me....
For the ML aficionados attending (real or virtual) 2021 MRS Fall Meeting -
@MaximZiatdinov
and yours truly will be giving tutorial CH04: Machine Learning and AI Methods for Materials Science—Applications to Imaging and Smart Experiments, 8:30 am – 5:00 pm ET, Monday, Nov. 29th.
A weird feelings for a scientist - reading a paper in a high profile journal that essentially repeats (idea, motivation, material) your decade old paper in another high profile journal. For a very small community. Any voice of advice,
#AcademicChatter
and
#AcademicTwitter
?
Bridging high resolution electron microscopy and supercomputing as integrated workflow. Kudos to
@MaximZiatdinov
of AtomAI, and non- twitterized Ayana Ghosh, as well as Bobby Sumpter and Ondrej Dyck.
Bayesian methods for active experiment - from simple BO and inference to deep kernel learning and structural GP and hypothesis learning. Can apply to other spaces - including chemical or processing.
@MaximZiatdinov
@ramav_matsci
@Yongtao_Liu
@GaryMarcus
@GaryMarcus
, I think you are now crossing the boundary. Disagreement with ideas is one that ng. Declaring person espousing different views as "the most dangerous person" is a totally different thing. Shame on you
Pleasant surprise of the morning - responding to professional and in depth reviews for two papers, where reviews clearly understood the concept and offered comments that will make paper clearer. Whomever you are, thank you!
This special feeling when your high school senior son tells you that he has written for fun LSTM predictor for SMILES -> solubility using AqSolDB data set
Machine learning for better hybrid perovskites -> Joule. Great collaboration and brainstorm with
@Robo_Perovskite
,
@MaximZiatdinov
, Eric Lass , and Yuanyuan Zhou.
Publication is sharing ideas. By now, we understand that sharing data and code is important. I wonder if publications should also have an optional share of materials (the samples may be available upon request) and training. Science is built as a community.
Lectures from ORNL workshop on Automated Experiment and Machine Learning in Scanning Probe Microscopy -> online. Featuring lectures by Mikhail Katsnelson , Leroy Cronin, Danilo J. Rezende , and tutorials on VAEs, DKL, DCNNs, and PyCroscopy. See:
@FunkEntropy
@timgill924
Many stand in awe at accomplishments of Tim Gill
@timgill924
, the world first quantum sociologist. But many still have to arrive to enlightenment.
1. Baby looks particularly contemplative after proving that the mass conservation laws in fact apply to babies.
2. Sproggy is confused - socks this small are difficult to chew
@BenBlaiszik
@andrewwhite01
I think for all of us part of the problem that our research is well outside classical domain sciences, yet not fully in the ML field (with NeurIPS, ICML, etc). This community is still very latent
The most amazing thing about all things automated experiment. One needs to know at some level:
- domain area (to even start)
- ML (by now beyond scikit-learn)
- elements of workflow building (Colab)
- IoT Integration
- and ideally cloud technology (AWS)
Sounds like a challenge!
Automated Scanning Probe Microscopy
@Yongtao_Liu
and hybrid perovskites
@Robo_Perovskite
can go hand in hand. Stand by for AE SPM that explores conductivity and current-voltage hysteresis (aka ionic motion) on grain boundaries only. There is some clear internal variability!
During my thesis defense exactly 20 years ago, half of my thesis committee (from materials science) argued that my thesis should be physics, and the second part (from physics) argued that it is Mat. Sci. Go figure - it is still not clear if I do physics/MSE or machine learning.
Our first publication on machine learning in STEM/EELS in 2010. We have been at it for a while - and it's great to see how well these predictions materialize (albeit it took deep kernel learning to do structure -EELS matching well).
They call it Windy City for a reason. On a positive side, I got to meet multiple denizens of the academic Twitterverse in AI for materials in person!
@taylordsparks
This is a fundamental problem with the ML community now - trying to learn domain sciences directly, rather than building a collaborative network with domain groups that have been invested in (simpler) ML for a while and have translational expertise
I love when ML researchers get excited about science, but seriously the reviewing process for scientific applications at ML conferences (e.g., ICLR) is entirely broken.
Papers with glaring errors are sailing through, without a single review from somebody with domain expertise.
One thing for sure - with ChatGPT, my students who had minimal or zero programming experience, now do homework problems in my course that 3 years ago would be a paper in high profile journal (in fact, that's where the problem has come me from).
@sama
and team did change the world
The first paper of the year with
@Robo_Perovskite
lab and building on the
@MaximZiatdinov
VAE libraries! For al ye hybrid perovskites and automated synthesis afficionados!
Speaking of review process - I wonder if it will be efficient to go from written reviews to scheduling an online anonymous Zoom session? There is no problem now to hide/alter voice. Can save time and make process fast and dynamic. If interesting - reshare!
#AcademicTwitter
For the ferroelectrics and PFM enthusiasts - now we have a capability to perform PFM spectroscopy when the temperature of the tip is dynamically changed. Read about it and contact Kyle Kelley to run on own samples via CNMS user program.
@fchollet
That's the first rule of using machine learning in experimental workflows: first do it without machine learning, and see what the bottlenecks are. May be they not even require machine learning!
The atomic construction community:
- half Russian-half Jewish
- Israeli
- Lebanese-Ukranian (Russian -speaking)
- Kurdish Iranian
We have the same goal!
Using Bayesian Optimization in the latent space of autoencoder trained on domain-specific examples. Here, we use it to create curly ferroelectrics - in theory so far. Will also work for materials and process optimization.
@MaximZiatdinov
@ramav_matsci
And now ELIT DCNN for STEM. Converting data stream into atomic positions in real time as a plug-in for NION SWIFT. We can find them, we can explore how they respond to beam.
@KevinRoccaprio1
@MaximZiatdinov