AntiFocal AI
@antifocal.bsky.social
Your scientific instruments tell you what happened. AntiFocal tells you what to try next: ranked experiments aimed at the open field, matched to the equipment your lab already has. For PIs, postdocs & core facilities.
🔬 Can an AI reading RHEED be checked more precisely than the XPS that labelled it?
Not against XPS alone. A neural network predicted the strontium fraction of 31 strontium titanate films from one RHEED image each, reaching r-squared 0.867 against XPS labels on 10 validation films.
Not against XPS alone. A neural network predicted the strontium fraction of 31 strontium titanate films from one RHEED image each, reaching r-squared 0.867 against XPS labels on 10 validation films.
October 28, 2026 at 9:25 AM
🔬 Can an AI reading RHEED be checked more precisely than the XPS that labelled it?
Not against XPS alone. A neural network predicted the strontium fraction of 31 strontium titanate films from one RHEED image each, reaching r-squared 0.867 against XPS labels on 10 validation films.
Not against XPS alone. A neural network predicted the strontium fraction of 31 strontium titanate films from one RHEED image each, reaching r-squared 0.867 against XPS labels on 10 validation films.
🔬 Does a falling prediction error mean the AI's recipes give better films?
In an MBE study of Fe₃Sn films, a random-forest model proposed 4 growths. Its error on a 0–100 film score spiked above 20, then fell to 9.9, which the authors call halving. Its leave-one-out error was about 10.
In an MBE study of Fe₃Sn films, a random-forest model proposed 4 growths. Its error on a 0–100 film score spiked above 20, then fell to 9.9, which the authors call halving. Its leave-one-out error was about 10.
October 27, 2026 at 3:01 PM
🔬 Does a falling prediction error mean the AI's recipes give better films?
In an MBE study of Fe₃Sn films, a random-forest model proposed 4 growths. Its error on a 0–100 film score spiked above 20, then fell to 9.9, which the authors call halving. Its leave-one-out error was about 10.
In an MBE study of Fe₃Sn films, a random-forest model proposed 4 growths. Its error on a 0–100 film score spiked above 20, then fell to 9.9, which the authors call halving. Its leave-one-out error was about 10.
Can a near-perfect fit tell you which parameters your data pin down? 🔬
Not by itself. In a WS₂ transistor study, on 1,000 simulated devices, two band-tail parameters erred in opposite directions in 77% of cases.
Not by itself. In a WS₂ transistor study, on 1,000 simulated devices, two band-tail parameters erred in opposite directions in 77% of cases.
October 27, 2026 at 12:06 PM
Can a near-perfect fit tell you which parameters your data pin down? 🔬
Not by itself. In a WS₂ transistor study, on 1,000 simulated devices, two band-tail parameters erred in opposite directions in 77% of cases.
Not by itself. In a WS₂ transistor study, on 1,000 simulated devices, two band-tail parameters erred in opposite directions in 77% of cases.
🔬 Can an AI agent align a crystal on a real synchrotron beamline?
In live tests, yes. A language-model agent found two reference reflections and the orientation matrix of a Co₃Sn₂S₂ crystal at 16.0 keV.
In live tests, yes. A language-model agent found two reference reflections and the orientation matrix of a Co₃Sn₂S₂ crystal at 16.0 keV.
October 26, 2026 at 8:18 AM
🔬 Can an AI agent align a crystal on a real synchrotron beamline?
In live tests, yes. A language-model agent found two reference reflections and the orientation matrix of a Co₃Sn₂S₂ crystal at 16.0 keV.
In live tests, yes. A language-model agent found two reference reflections and the orientation matrix of a Co₃Sn₂S₂ crystal at 16.0 keV.
🔬 𝗗𝗼𝗲𝘀 𝗮𝗻 𝗔𝗜 𝗻𝗲𝗲𝗱 𝘁𝗼 𝗸𝗻𝗼𝘄 𝘄𝗵𝗮𝘁 𝗮𝗻 𝗮𝗯𝘀𝗼𝗿𝗽𝘁𝗶𝗼𝗻 𝗲𝗱𝗴𝗲 𝗹𝗼𝗼𝗸𝘀 𝗹𝗶𝗸𝗲 𝘁𝗼 𝗺𝗲𝗮𝘀𝘂𝗿𝗲 𝗼𝗻𝗲?
In one 2025 study it mattered. At Argonne's Advanced Photon Source, an edge-aware AI tracked the nickel edge of a working battery with 60 points per spectrum.
In one 2025 study it mattered. At Argonne's Advanced Photon Source, an edge-aware AI tracked the nickel edge of a working battery with 60 points per spectrum.
October 24, 2026 at 5:34 PM
🔬 𝗗𝗼𝗲𝘀 𝗮𝗻 𝗔𝗜 𝗻𝗲𝗲𝗱 𝘁𝗼 𝗸𝗻𝗼𝘄 𝘄𝗵𝗮𝘁 𝗮𝗻 𝗮𝗯𝘀𝗼𝗿𝗽𝘁𝗶𝗼𝗻 𝗲𝗱𝗴𝗲 𝗹𝗼𝗼𝗸𝘀 𝗹𝗶𝗸𝗲 𝘁𝗼 𝗺𝗲𝗮𝘀𝘂𝗿𝗲 𝗼𝗻𝗲?
In one 2025 study it mattered. At Argonne's Advanced Photon Source, an edge-aware AI tracked the nickel edge of a working battery with 60 points per spectrum.
In one 2025 study it mattered. At Argonne's Advanced Photon Source, an edge-aware AI tracked the nickel edge of a working battery with 60 points per spectrum.
🔬 𝗗𝗼𝗲𝘀 𝗮𝗻 𝗔𝗜 𝘀𝘁𝗲𝗲𝗿𝗶𝗻𝗴 𝗮 𝗺𝗶𝗰𝗿𝗼𝗔𝗥𝗣𝗘𝗦 𝗯𝗲𝗮𝗺 𝗱𝗲𝗰𝗶𝗱𝗲 𝘄𝗵𝗮𝘁 𝗶𝘀 𝗶𝗻𝘁𝗲𝗿𝗲𝘀𝘁𝗶𝗻𝗴 𝗼𝗻 𝘁𝗵𝗲 𝘀𝗮𝗺𝗽𝗹𝗲?
No. It picks where to look; people define interesting. By 300 points its map showed fine structure; the reference grid had about 22,700.
No. It picks where to look; people define interesting. By 300 points its map showed fine structure; the reference grid had about 22,700.
October 24, 2026 at 3:59 PM
🔬 𝗗𝗼𝗲𝘀 𝗮𝗻 𝗔𝗜 𝘀𝘁𝗲𝗲𝗿𝗶𝗻𝗴 𝗮 𝗺𝗶𝗰𝗿𝗼𝗔𝗥𝗣𝗘𝗦 𝗯𝗲𝗮𝗺 𝗱𝗲𝗰𝗶𝗱𝗲 𝘄𝗵𝗮𝘁 𝗶𝘀 𝗶𝗻𝘁𝗲𝗿𝗲𝘀𝘁𝗶𝗻𝗴 𝗼𝗻 𝘁𝗵𝗲 𝘀𝗮𝗺𝗽𝗹𝗲?
No. It picks where to look; people define interesting. By 300 points its map showed fine structure; the reference grid had about 22,700.
No. It picks where to look; people define interesting. By 300 points its map showed fine structure; the reference grid had about 22,700.
🔬 𝗦𝗵𝗼𝘂𝗹𝗱 𝗮𝗻 𝗔𝗜 𝗶𝗻𝗱𝗲𝗻𝘁 𝘄𝗵𝗲𝗿𝗲 𝗶𝘁 𝗸𝗻𝗼𝘄𝘀 𝗹𝗲𝗮𝘀𝘁?
On a sputtered Ta-Ti-Hf-Zr wafer, a nanoindenter loop that priced every move mapped hardness in 43,176 s. The same loop without the cost model was reported at 150,780 s to the same endpoint, about 3.5× slower.
On a sputtered Ta-Ti-Hf-Zr wafer, a nanoindenter loop that priced every move mapped hardness in 43,176 s. The same loop without the cost model was reported at 150,780 s to the same endpoint, about 3.5× slower.
October 24, 2026 at 1:59 PM
🔬 𝗦𝗵𝗼𝘂𝗹𝗱 𝗮𝗻 𝗔𝗜 𝗶𝗻𝗱𝗲𝗻𝘁 𝘄𝗵𝗲𝗿𝗲 𝗶𝘁 𝗸𝗻𝗼𝘄𝘀 𝗹𝗲𝗮𝘀𝘁?
On a sputtered Ta-Ti-Hf-Zr wafer, a nanoindenter loop that priced every move mapped hardness in 43,176 s. The same loop without the cost model was reported at 150,780 s to the same endpoint, about 3.5× slower.
On a sputtered Ta-Ti-Hf-Zr wafer, a nanoindenter loop that priced every move mapped hardness in 43,176 s. The same loop without the cost model was reported at 150,780 s to the same endpoint, about 3.5× slower.
🔬 𝗖𝗮𝗻 𝗔𝗜 𝗿𝗲𝗮𝗱 𝘁𝗵𝗲 𝗺𝗲𝗰𝗵𝗮𝗻𝗶𝘀𝗺 𝗳𝗿𝗼𝗺 𝗼𝗻𝗲 𝘃𝗼𝗹𝘁𝗮𝗺𝗺𝗼𝗴𝗿𝗮𝗺?
Not from one scan. A 2024 deep-learning detector boxes each redox event in a cyclic voltammogram and ranks 8 mechanisms. On simulated data it named up to 97.2% of the events it found correctly.
Not from one scan. A 2024 deep-learning detector boxes each redox event in a cyclic voltammogram and ranks 8 mechanisms. On simulated data it named up to 97.2% of the events it found correctly.
October 24, 2026 at 12:03 PM
🔬 𝗖𝗮𝗻 𝗔𝗜 𝗿𝗲𝗮𝗱 𝘁𝗵𝗲 𝗺𝗲𝗰𝗵𝗮𝗻𝗶𝘀𝗺 𝗳𝗿𝗼𝗺 𝗼𝗻𝗲 𝘃𝗼𝗹𝘁𝗮𝗺𝗺𝗼𝗴𝗿𝗮𝗺?
Not from one scan. A 2024 deep-learning detector boxes each redox event in a cyclic voltammogram and ranks 8 mechanisms. On simulated data it named up to 97.2% of the events it found correctly.
Not from one scan. A 2024 deep-learning detector boxes each redox event in a cyclic voltammogram and ranks 8 mechanisms. On simulated data it named up to 97.2% of the events it found correctly.
🔬 𝗖𝗮𝗻 𝗔𝗜 𝗳𝗶𝗻𝗱 𝘁𝗵𝗲 𝗿𝗲𝗰𝗶𝗽𝗲 𝗶𝗳 𝘆𝗼𝘂 𝗼𝗻𝗹𝘆 𝗴𝗶𝘃𝗲 𝗶𝘁 𝘁𝗵𝗲 𝘁𝗮𝗿𝗴𝗲𝘁 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲?
Partly. In a 2025 preprint a robot at the DanMAX synchrotron beamline got close to a 5 nm gold decahedron target in 41 runs.
Partly. In a 2025 preprint a robot at the DanMAX synchrotron beamline got close to a 5 nm gold decahedron target in 41 runs.
October 23, 2026 at 10:24 PM
🔬 𝗖𝗮𝗻 𝗔𝗜 𝗳𝗶𝗻𝗱 𝘁𝗵𝗲 𝗿𝗲𝗰𝗶𝗽𝗲 𝗶𝗳 𝘆𝗼𝘂 𝗼𝗻𝗹𝘆 𝗴𝗶𝘃𝗲 𝗶𝘁 𝘁𝗵𝗲 𝘁𝗮𝗿𝗴𝗲𝘁 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲?
Partly. In a 2025 preprint a robot at the DanMAX synchrotron beamline got close to a 5 nm gold decahedron target in 41 runs.
Partly. In a 2025 preprint a robot at the DanMAX synchrotron beamline got close to a 5 nm gold decahedron target in 41 runs.
🔬 𝗗𝗼𝗲𝘀 𝗮𝗻 𝗔𝗜 𝗻𝗲𝗲𝗱 𝗦𝗘𝗠 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝘁𝗼 𝗰𝗵𝗼𝗼𝘀𝗲 𝘄𝗵𝗲𝗿𝗲 𝘁𝗵𝗲 𝗺𝗶𝗰𝗿𝗼𝘀𝗰𝗼𝗽𝗲 𝘇𝗼𝗼𝗺𝘀 𝗶𝗻?
Not in one new University of Toronto preprint. A general vision model with no SEM fine-tuning steered a field-emission SEM, which acquired 78 micrographs automatically at 500×, 3,500× and 24,500×.
Not in one new University of Toronto preprint. A general vision model with no SEM fine-tuning steered a field-emission SEM, which acquired 78 micrographs automatically at 500×, 3,500× and 24,500×.
October 19, 2026 at 2:12 PM
🔬 𝗗𝗼𝗲𝘀 𝗮𝗻 𝗔𝗜 𝗻𝗲𝗲𝗱 𝗦𝗘𝗠 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝘁𝗼 𝗰𝗵𝗼𝗼𝘀𝗲 𝘄𝗵𝗲𝗿𝗲 𝘁𝗵𝗲 𝗺𝗶𝗰𝗿𝗼𝘀𝗰𝗼𝗽𝗲 𝘇𝗼𝗼𝗺𝘀 𝗶𝗻?
Not in one new University of Toronto preprint. A general vision model with no SEM fine-tuning steered a field-emission SEM, which acquired 78 micrographs automatically at 500×, 3,500× and 24,500×.
Not in one new University of Toronto preprint. A general vision model with no SEM fine-tuning steered a field-emission SEM, which acquired 78 micrographs automatically at 500×, 3,500× and 24,500×.
🔬 𝗗𝗼𝗲𝘀 𝗮 𝘀𝗲𝗹𝗳-𝗱𝗿𝗶𝘃𝗶𝗻𝗴 𝗹𝗮𝗯 𝗻𝗲𝗲𝗱 𝗼𝗻𝗲 𝘀𝗮𝗺𝗽𝗹𝗲 𝗽𝗲𝗿 𝗲𝘅𝗽𝗲𝗿𝗶𝗺𝗲𝗻𝘁?
At NIMS in Japan, an AI loop searched sputtered alloy films for a large anomalous Hall effect. Each film carried 13 compositions. In 18 cycles it measured 234; the top film reached 10.9 µΩ cm in cycle 11.
At NIMS in Japan, an AI loop searched sputtered alloy films for a large anomalous Hall effect. Each film carried 13 compositions. In 18 cycles it measured 234; the top film reached 10.9 µΩ cm in cycle 11.
October 19, 2026 at 2:11 PM
🔬 𝗗𝗼𝗲𝘀 𝗮 𝘀𝗲𝗹𝗳-𝗱𝗿𝗶𝘃𝗶𝗻𝗴 𝗹𝗮𝗯 𝗻𝗲𝗲𝗱 𝗼𝗻𝗲 𝘀𝗮𝗺𝗽𝗹𝗲 𝗽𝗲𝗿 𝗲𝘅𝗽𝗲𝗿𝗶𝗺𝗲𝗻𝘁?
At NIMS in Japan, an AI loop searched sputtered alloy films for a large anomalous Hall effect. Each film carried 13 compositions. In 18 cycles it measured 234; the top film reached 10.9 µΩ cm in cycle 11.
At NIMS in Japan, an AI loop searched sputtered alloy films for a large anomalous Hall effect. Each film carried 13 compositions. In 18 cycles it measured 234; the top film reached 10.9 µΩ cm in cycle 11.
🔬 𝗦𝗵𝗼𝘂𝗹𝗱 𝗮 𝘀𝗲𝗹𝗳-𝗱𝗿𝗶𝘃𝗶𝗻𝗴 𝗹𝗮𝗯 𝗵𝗮𝗻𝗱 𝗯𝗮𝗰𝗸 𝗼𝗻𝗹𝘆 𝗶𝘁𝘀 𝗯𝗲𝘀𝘁 𝗿𝗲𝗰𝗶𝗽𝗲?
At NTT, Bayesian optimisation ran 66 sputtered gallium oxide films and reached an Urbach energy of 182 meV at run 56. A random forest turned the runs into a growth rule. NTT reports a researcher following it reached 163 meV.
At NTT, Bayesian optimisation ran 66 sputtered gallium oxide films and reached an Urbach energy of 182 meV at run 56. A random forest turned the runs into a growth rule. NTT reports a researcher following it reached 163 meV.
October 19, 2026 at 2:10 PM
🔬 𝗦𝗵𝗼𝘂𝗹𝗱 𝗮 𝘀𝗲𝗹𝗳-𝗱𝗿𝗶𝘃𝗶𝗻𝗴 𝗹𝗮𝗯 𝗵𝗮𝗻𝗱 𝗯𝗮𝗰𝗸 𝗼𝗻𝗹𝘆 𝗶𝘁𝘀 𝗯𝗲𝘀𝘁 𝗿𝗲𝗰𝗶𝗽𝗲?
At NTT, Bayesian optimisation ran 66 sputtered gallium oxide films and reached an Urbach energy of 182 meV at run 56. A random forest turned the runs into a growth rule. NTT reports a researcher following it reached 163 meV.
At NTT, Bayesian optimisation ran 66 sputtered gallium oxide films and reached an Urbach energy of 182 meV at run 56. A random forest turned the runs into a growth rule. NTT reports a researcher following it reached 163 meV.
🔬 𝗖𝗮𝗻 𝗮𝗻 𝗔𝗜 𝘀𝗲𝗲 𝗮𝘁𝗼𝗺𝗶𝗰 𝗼𝗿𝗱𝗲𝗿 𝘁𝗵𝗮𝘁 𝗮𝗻 𝗮𝘁𝗼𝗺 𝗽𝗿𝗼𝗯𝗲 𝗰𝗼𝗺𝗽𝗼𝘀𝗶𝘁𝗶𝗼𝗻 𝗺𝗮𝗽 𝗺𝗶𝘀𝘀𝗲𝘀?
Partly. A 3D network trained only on simulated lattices found signs of copper–gold ordering in red gold that the usual composition isosurface cannot separate.
Partly. A 3D network trained only on simulated lattices found signs of copper–gold ordering in red gold that the usual composition isosurface cannot separate.
October 19, 2026 at 2:09 PM
🔬 𝗖𝗮𝗻 𝗮𝗻 𝗔𝗜 𝘀𝗲𝗲 𝗮𝘁𝗼𝗺𝗶𝗰 𝗼𝗿𝗱𝗲𝗿 𝘁𝗵𝗮𝘁 𝗮𝗻 𝗮𝘁𝗼𝗺 𝗽𝗿𝗼𝗯𝗲 𝗰𝗼𝗺𝗽𝗼𝘀𝗶𝘁𝗶𝗼𝗻 𝗺𝗮𝗽 𝗺𝗶𝘀𝘀𝗲𝘀?
Partly. A 3D network trained only on simulated lattices found signs of copper–gold ordering in red gold that the usual composition isosurface cannot separate.
Partly. A 3D network trained only on simulated lattices found signs of copper–gold ordering in red gold that the usual composition isosurface cannot separate.
🔬 𝗗𝗼𝗲𝘀 𝗮𝗻 𝗔𝗜 𝘁𝗵𝗮𝘁 𝗱𝗲𝘃𝗲𝗹𝗼𝗽𝘀 𝗛𝗣𝗟𝗖 𝗺𝗲𝘁𝗵𝗼𝗱𝘀 𝗼𝘃𝗲𝗿𝗻𝗶𝗴𝗵𝘁 𝗱𝗲𝗰𝗶𝗱𝗲 𝘄𝗵𝗮𝘁 𝗮 𝗴𝗼𝗼𝗱 𝘀𝗲𝗽𝗮𝗿𝗮𝘁𝗶𝗼𝗻 𝗶𝘀?
No. A University of Leeds team let Bayesian optimisation run an HPLC overnight. It separated seven compounds at resolution 2 or more in about 13 experiments on average, to a score the researchers wrote.
No. A University of Leeds team let Bayesian optimisation run an HPLC overnight. It separated seven compounds at resolution 2 or more in about 13 experiments on average, to a score the researchers wrote.
October 19, 2026 at 2:08 PM
🔬 𝗗𝗼𝗲𝘀 𝗮𝗻 𝗔𝗜 𝘁𝗵𝗮𝘁 𝗱𝗲𝘃𝗲𝗹𝗼𝗽𝘀 𝗛𝗣𝗟𝗖 𝗺𝗲𝘁𝗵𝗼𝗱𝘀 𝗼𝘃𝗲𝗿𝗻𝗶𝗴𝗵𝘁 𝗱𝗲𝗰𝗶𝗱𝗲 𝘄𝗵𝗮𝘁 𝗮 𝗴𝗼𝗼𝗱 𝘀𝗲𝗽𝗮𝗿𝗮𝘁𝗶𝗼𝗻 𝗶𝘀?
No. A University of Leeds team let Bayesian optimisation run an HPLC overnight. It separated seven compounds at resolution 2 or more in about 13 experiments on average, to a score the researchers wrote.
No. A University of Leeds team let Bayesian optimisation run an HPLC overnight. It separated seven compounds at resolution 2 or more in about 13 experiments on average, to a score the researchers wrote.
🔬 𝗗𝗼𝗲𝘀 𝗮𝗻 𝗔𝗜 𝘁𝗵𝗮𝘁 𝗴𝗿𝗼𝘄𝘀 𝗴𝗿𝗮𝗽𝗵𝗲𝗻𝗲 𝗻𝗲𝗲𝗱 𝗮 𝗹𝗶𝗯𝗿𝗮𝗿𝘆 𝗼𝗳 𝗼𝗹𝗱 𝗿𝗲𝗰𝗶𝗽𝗲𝘀?
It started with none. A neural network proposed heating profiles for graphene on silicon carbide and learned from one Raman spectrum per run. Meaningful learning came within a few tens of runs.
It started with none. A neural network proposed heating profiles for graphene on silicon carbide and learned from one Raman spectrum per run. Meaningful learning came within a few tens of runs.
October 14, 2026 at 9:55 PM
🔬 𝗗𝗼𝗲𝘀 𝗮𝗻 𝗔𝗜 𝘁𝗵𝗮𝘁 𝗴𝗿𝗼𝘄𝘀 𝗴𝗿𝗮𝗽𝗵𝗲𝗻𝗲 𝗻𝗲𝗲𝗱 𝗮 𝗹𝗶𝗯𝗿𝗮𝗿𝘆 𝗼𝗳 𝗼𝗹𝗱 𝗿𝗲𝗰𝗶𝗽𝗲𝘀?
It started with none. A neural network proposed heating profiles for graphene on silicon carbide and learned from one Raman spectrum per run. Meaningful learning came within a few tens of runs.
It started with none. A neural network proposed heating profiles for graphene on silicon carbide and learned from one Raman spectrum per run. Meaningful learning came within a few tens of runs.
🔬 𝗜𝘀 𝗔𝗜 𝗼𝗻 𝗮𝗻 𝗲𝗹𝗲𝗰𝘁𝗿𝗼𝗻 𝗺𝗶𝗰𝗿𝗼𝘀𝗰𝗼𝗽𝗲 𝗼𝗻𝗹𝘆 𝘁𝗵𝗲𝗿𝗲 𝘁𝗼 𝗰𝗹𝗲𝗮𝗻 𝘂𝗽 𝘁𝗵𝗲 𝗶𝗺𝗮𝗴𝗲?
In a 2024 STEM-EELS study, neural networks found defects in each fast image of vanadium-doped MoS₂. Software picked about 10 atomic sites per image pair for spectra. One vanadium atom showed its L-edge in a 1 s spectrum.
In a 2024 STEM-EELS study, neural networks found defects in each fast image of vanadium-doped MoS₂. Software picked about 10 atomic sites per image pair for spectra. One vanadium atom showed its L-edge in a 1 s spectrum.
October 14, 2026 at 9:54 PM
🔬 𝗜𝘀 𝗔𝗜 𝗼𝗻 𝗮𝗻 𝗲𝗹𝗲𝗰𝘁𝗿𝗼𝗻 𝗺𝗶𝗰𝗿𝗼𝘀𝗰𝗼𝗽𝗲 𝗼𝗻𝗹𝘆 𝘁𝗵𝗲𝗿𝗲 𝘁𝗼 𝗰𝗹𝗲𝗮𝗻 𝘂𝗽 𝘁𝗵𝗲 𝗶𝗺𝗮𝗴𝗲?
In a 2024 STEM-EELS study, neural networks found defects in each fast image of vanadium-doped MoS₂. Software picked about 10 atomic sites per image pair for spectra. One vanadium atom showed its L-edge in a 1 s spectrum.
In a 2024 STEM-EELS study, neural networks found defects in each fast image of vanadium-doped MoS₂. Software picked about 10 atomic sites per image pair for spectra. One vanadium atom showed its L-edge in a 1 s spectrum.
🔬 𝗖𝗮𝗻 𝗮 𝗻𝗲𝘂𝗿𝗮𝗹 𝗻𝗲𝘁𝘄𝗼𝗿𝗸 𝗳𝗶𝘅 𝗮 𝗯𝗮𝗱𝗹𝘆 𝘀𝗵𝗶𝗺𝗺𝗲𝗱 𝗡𝗠𝗥 𝘀𝗽𝗲𝗰𝘁𝗿𝘂𝗺 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗸𝗻𝗼𝘄𝗶𝗻𝗴 𝘁𝗵𝗲 𝗺𝗮𝗴𝗻𝗲𝘁?
Not on its own. A University of Warsaw group first had a 700 MHz spectrometer record 10,201 deliberately mis-shimmed spectra over about 33 hours. The network learned that magnet's own line distortions.
Not on its own. A University of Warsaw group first had a 700 MHz spectrometer record 10,201 deliberately mis-shimmed spectra over about 33 hours. The network learned that magnet's own line distortions.
October 14, 2026 at 9:53 PM
🔬 𝗖𝗮𝗻 𝗮 𝗻𝗲𝘂𝗿𝗮𝗹 𝗻𝗲𝘁𝘄𝗼𝗿𝗸 𝗳𝗶𝘅 𝗮 𝗯𝗮𝗱𝗹𝘆 𝘀𝗵𝗶𝗺𝗺𝗲𝗱 𝗡𝗠𝗥 𝘀𝗽𝗲𝗰𝘁𝗿𝘂𝗺 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗸𝗻𝗼𝘄𝗶𝗻𝗴 𝘁𝗵𝗲 𝗺𝗮𝗴𝗻𝗲𝘁?
Not on its own. A University of Warsaw group first had a 700 MHz spectrometer record 10,201 deliberately mis-shimmed spectra over about 33 hours. The network learned that magnet's own line distortions.
Not on its own. A University of Warsaw group first had a 700 MHz spectrometer record 10,201 deliberately mis-shimmed spectra over about 33 hours. The network learned that magnet's own line distortions.
🔬 𝗖𝗮𝗻 𝗮𝗻 𝗔𝗜 𝘁𝘂𝗻𝗲 𝗮 𝗖𝗩𝗗 𝗳𝘂𝗿𝗻𝗮𝗰𝗲 𝗯𝘆 𝘄𝗮𝘁𝗰𝗵𝗶𝗻𝗴 𝘁𝗵𝗲 𝗳𝗶𝗹𝗺 𝗴𝗹𝗼𝘄?
Up to a point. At NIMS in Japan, Bayesian optimisation tuned four CVD settings for monolayer WS₂. By round 7 of 13 it found conditions giving a film 86.6% brighter in photoluminescence than the initial data.
Up to a point. At NIMS in Japan, Bayesian optimisation tuned four CVD settings for monolayer WS₂. By round 7 of 13 it found conditions giving a film 86.6% brighter in photoluminescence than the initial data.
October 14, 2026 at 9:52 PM
🔬 𝗖𝗮𝗻 𝗮𝗻 𝗔𝗜 𝘁𝘂𝗻𝗲 𝗮 𝗖𝗩𝗗 𝗳𝘂𝗿𝗻𝗮𝗰𝗲 𝗯𝘆 𝘄𝗮𝘁𝗰𝗵𝗶𝗻𝗴 𝘁𝗵𝗲 𝗳𝗶𝗹𝗺 𝗴𝗹𝗼𝘄?
Up to a point. At NIMS in Japan, Bayesian optimisation tuned four CVD settings for monolayer WS₂. By round 7 of 13 it found conditions giving a film 86.6% brighter in photoluminescence than the initial data.
Up to a point. At NIMS in Japan, Bayesian optimisation tuned four CVD settings for monolayer WS₂. By round 7 of 13 it found conditions giving a film 86.6% brighter in photoluminescence than the initial data.
🔬 𝗖𝗮𝗻 𝗮 𝗿𝗼𝗯𝗼𝘁 𝗹𝗲𝗮𝗿𝗻 𝗵𝗼𝘄 𝗺𝘂𝗰𝗵 𝗵𝘂𝗺𝗶𝗱𝗶𝘁𝘆 𝗮 𝗽𝗲𝗿𝗼𝘃𝘀𝗸𝗶𝘁𝗲 𝗳𝗶𝗹𝗺 𝘁𝗼𝗹𝗲𝗿𝗮𝘁𝗲𝘀 𝗳𝗿𝗼𝗺 𝟰𝟬 𝗿𝗲𝗰𝗶𝗽𝗲𝘀?
A Berkeley Lab robot tested 40 of 5,000+ recipe combinations, 160 films in all. Its model predicted scores above 80 at 5–25% humidity and 66 or lower from 35% up.
A Berkeley Lab robot tested 40 of 5,000+ recipe combinations, 160 films in all. Its model predicted scores above 80 at 5–25% humidity and 66 or lower from 35% up.
October 10, 2026 at 7:45 PM
🔬 𝗖𝗮𝗻 𝗮 𝗿𝗼𝗯𝗼𝘁 𝗹𝗲𝗮𝗿𝗻 𝗵𝗼𝘄 𝗺𝘂𝗰𝗵 𝗵𝘂𝗺𝗶𝗱𝗶𝘁𝘆 𝗮 𝗽𝗲𝗿𝗼𝘃𝘀𝗸𝗶𝘁𝗲 𝗳𝗶𝗹𝗺 𝘁𝗼𝗹𝗲𝗿𝗮𝘁𝗲𝘀 𝗳𝗿𝗼𝗺 𝟰𝟬 𝗿𝗲𝗰𝗶𝗽𝗲𝘀?
A Berkeley Lab robot tested 40 of 5,000+ recipe combinations, 160 films in all. Its model predicted scores above 80 at 5–25% humidity and 66 or lower from 35% up.
A Berkeley Lab robot tested 40 of 5,000+ recipe combinations, 160 films in all. Its model predicted scores above 80 at 5–25% humidity and 66 or lower from 35% up.
🔬 𝗖𝗮𝗻 𝗮 𝗻𝗲𝘂𝗿𝗮𝗹 𝗻𝗲𝘁𝘄𝗼𝗿𝗸 𝗺𝗲𝗮𝘀𝘂𝗿𝗲 𝗮 𝗻𝗮𝗻𝗼𝗴𝗿𝗮𝘁𝗶𝗻𝗴 𝘁𝗼 𝟬.𝟭 𝗻𝗺 𝗳𝗿𝗼𝗺 𝗲𝗹𝗹𝗶𝗽𝘀𝗼𝗺𝗲𝘁𝗿𝘆 𝗱𝗮𝘁𝗮?
A neural network turned ellipsometry Mueller matrices into grating height and width in 132 ms, against 7.3 hours for iterative fitting. Average error: 0.103 nm. But all 20 test spectra were simulated.
A neural network turned ellipsometry Mueller matrices into grating height and width in 132 ms, against 7.3 hours for iterative fitting. Average error: 0.103 nm. But all 20 test spectra were simulated.
October 9, 2026 at 9:06 PM
🔬 𝗖𝗮𝗻 𝗮 𝗻𝗲𝘂𝗿𝗮𝗹 𝗻𝗲𝘁𝘄𝗼𝗿𝗸 𝗺𝗲𝗮𝘀𝘂𝗿𝗲 𝗮 𝗻𝗮𝗻𝗼𝗴𝗿𝗮𝘁𝗶𝗻𝗴 𝘁𝗼 𝟬.𝟭 𝗻𝗺 𝗳𝗿𝗼𝗺 𝗲𝗹𝗹𝗶𝗽𝘀𝗼𝗺𝗲𝘁𝗿𝘆 𝗱𝗮𝘁𝗮?
A neural network turned ellipsometry Mueller matrices into grating height and width in 132 ms, against 7.3 hours for iterative fitting. Average error: 0.103 nm. But all 20 test spectra were simulated.
A neural network turned ellipsometry Mueller matrices into grating height and width in 132 ms, against 7.3 hours for iterative fitting. Average error: 0.103 nm. But all 20 test spectra were simulated.
🔬 𝗗𝗼𝗲𝘀 𝗮𝗻 𝗔𝗜 𝗳𝗶𝘁 𝘁𝗼 𝗮𝗻 𝗫𝗥𝗗 𝗽𝗮𝘁𝘁𝗲𝗿𝗻 𝗽𝗿𝗼𝘃𝗲 𝘆𝗼𝘂 𝗺𝗮𝗱𝗲 𝗮 𝗻𝗲𝘄 𝗺𝗮𝘁𝗲𝗿𝗶𝗮𝗹?
In 2023 an autonomous lab reported 41 new compounds from 58 targets in 17 days. AI read the powder X-ray diffraction of every product.
In 2023 an autonomous lab reported 41 new compounds from 58 targets in 17 days. AI read the powder X-ray diffraction of every product.
October 7, 2026 at 4:21 PM
🔬 𝗗𝗼𝗲𝘀 𝗮𝗻 𝗔𝗜 𝗳𝗶𝘁 𝘁𝗼 𝗮𝗻 𝗫𝗥𝗗 𝗽𝗮𝘁𝘁𝗲𝗿𝗻 𝗽𝗿𝗼𝘃𝗲 𝘆𝗼𝘂 𝗺𝗮𝗱𝗲 𝗮 𝗻𝗲𝘄 𝗺𝗮𝘁𝗲𝗿𝗶𝗮𝗹?
In 2023 an autonomous lab reported 41 new compounds from 58 targets in 17 days. AI read the powder X-ray diffraction of every product.
In 2023 an autonomous lab reported 41 new compounds from 58 targets in 17 days. AI read the powder X-ray diffraction of every product.
🔬 𝗗𝗼𝗲𝘀 𝗮 𝗰𝗼𝗮𝘁𝗶𝗻𝗴 𝗿𝗲𝗰𝗶𝗽𝗲 𝗮𝗻 𝗔𝗜 𝘁𝘂𝗻𝗲𝘀 𝗼𝗻 𝟮 𝗰𝗺 𝗰𝗵𝗶𝗽𝘀 𝗵𝗼𝗹𝗱 𝗼𝗻 𝗮 𝟭𝟬𝟬 𝗰𝗺² 𝗳𝗶𝗹𝗺?
In a 2025 study a robot coated a conducting polymer on 2 cm chips as an AI picked each next recipe. A recipe from that search gave over 4,500 S/cm on a 100 cm² film once people chose it and added two rinses.
In a 2025 study a robot coated a conducting polymer on 2 cm chips as an AI picked each next recipe. A recipe from that search gave over 4,500 S/cm on a 100 cm² film once people chose it and added two rinses.
October 7, 2026 at 4:19 PM
🔬 𝗗𝗼𝗲𝘀 𝗮 𝗰𝗼𝗮𝘁𝗶𝗻𝗴 𝗿𝗲𝗰𝗶𝗽𝗲 𝗮𝗻 𝗔𝗜 𝘁𝘂𝗻𝗲𝘀 𝗼𝗻 𝟮 𝗰𝗺 𝗰𝗵𝗶𝗽𝘀 𝗵𝗼𝗹𝗱 𝗼𝗻 𝗮 𝟭𝟬𝟬 𝗰𝗺² 𝗳𝗶𝗹𝗺?
In a 2025 study a robot coated a conducting polymer on 2 cm chips as an AI picked each next recipe. A recipe from that search gave over 4,500 S/cm on a 100 cm² film once people chose it and added two rinses.
In a 2025 study a robot coated a conducting polymer on 2 cm chips as an AI picked each next recipe. A recipe from that search gave over 4,500 S/cm on a 100 cm² film once people chose it and added two rinses.
🔬 𝗗𝗼𝗲𝘀 𝗮𝗻 𝗔𝗜 𝗳𝗶𝘁 𝘁𝗼 𝗮𝗻 𝗫𝗥𝗗 𝗽𝗮𝘁𝘁𝗲𝗿𝗻 𝗽𝗿𝗼𝘃𝗲 𝘆𝗼𝘂 𝗺𝗮𝗱𝗲 𝗮 𝗻𝗲𝘄 𝗺𝗮𝘁𝗲𝗿𝗶𝗮𝗹?
In 2023 an autonomous lab reported 41 new compounds from 58 targets in 17 days. AI read the powder X-ray diffraction of every product.
In 2023 an autonomous lab reported 41 new compounds from 58 targets in 17 days. AI read the powder X-ray diffraction of every product.
October 5, 2026 at 12:40 PM
🔬 𝗗𝗼𝗲𝘀 𝗮𝗻 𝗔𝗜 𝗳𝗶𝘁 𝘁𝗼 𝗮𝗻 𝗫𝗥𝗗 𝗽𝗮𝘁𝘁𝗲𝗿𝗻 𝗽𝗿𝗼𝘃𝗲 𝘆𝗼𝘂 𝗺𝗮𝗱𝗲 𝗮 𝗻𝗲𝘄 𝗺𝗮𝘁𝗲𝗿𝗶𝗮𝗹?
In 2023 an autonomous lab reported 41 new compounds from 58 targets in 17 days. AI read the powder X-ray diffraction of every product.
In 2023 an autonomous lab reported 41 new compounds from 58 targets in 17 days. AI read the powder X-ray diffraction of every product.
🔬 𝗜𝘀 𝗮𝗻 𝗔𝗜 𝗳𝗮𝘀𝘁𝗲𝗿 𝘁𝗵𝗮𝗻 𝗮 𝗽𝗵𝘆𝘀𝗶𝗰𝗶𝘀𝘁 𝗮𝘁 𝘁𝘂𝗻𝗶𝗻𝗴 𝗮 𝗾𝘂𝗮𝗻𝘁𝘂𝗺 𝗱𝗼𝘁 𝗱𝗲𝘃𝗶𝗰𝗲?
In one study, a tuning algorithm brought GaAs double quantum dots at 50 mK into a double-dot regime in a median 0.6 to 2.8 h, depending on the device and setup. The authors put experienced people at about 3 h.
In one study, a tuning algorithm brought GaAs double quantum dots at 50 mK into a double-dot regime in a median 0.6 to 2.8 h, depending on the device and setup. The authors put experienced people at about 3 h.
October 4, 2026 at 7:36 PM
🔬 𝗜𝘀 𝗮𝗻 𝗔𝗜 𝗳𝗮𝘀𝘁𝗲𝗿 𝘁𝗵𝗮𝗻 𝗮 𝗽𝗵𝘆𝘀𝗶𝗰𝗶𝘀𝘁 𝗮𝘁 𝘁𝘂𝗻𝗶𝗻𝗴 𝗮 𝗾𝘂𝗮𝗻𝘁𝘂𝗺 𝗱𝗼𝘁 𝗱𝗲𝘃𝗶𝗰𝗲?
In one study, a tuning algorithm brought GaAs double quantum dots at 50 mK into a double-dot regime in a median 0.6 to 2.8 h, depending on the device and setup. The authors put experienced people at about 3 h.
In one study, a tuning algorithm brought GaAs double quantum dots at 50 mK into a double-dot regime in a median 0.6 to 2.8 h, depending on the device and setup. The authors put experienced people at about 3 h.
🔬 𝗖𝗮𝗻 𝗮𝗻 𝗔𝗜 𝗻𝗮𝗺𝗲 𝗮 𝗰𝗿𝘆𝘀𝘁𝗮𝗹 𝗳𝗿𝗼𝗺 𝗶𝘁𝘀 𝗘𝗕𝗦𝗗 𝗽𝗮𝘁𝘁𝗲𝗿𝗻 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗮 𝗽𝗵𝗮𝘀𝗲 𝗹𝗶𝘀𝘁?
A convolutional neural network from UC San Diego reads crystal symmetry straight from an EBSD Kikuchi pattern, with no candidate phase list. The university reported at least 95% accuracy.
A convolutional neural network from UC San Diego reads crystal symmetry straight from an EBSD Kikuchi pattern, with no candidate phase list. The university reported at least 95% accuracy.
October 3, 2026 at 1:57 PM
🔬 𝗖𝗮𝗻 𝗮𝗻 𝗔𝗜 𝗻𝗮𝗺𝗲 𝗮 𝗰𝗿𝘆𝘀𝘁𝗮𝗹 𝗳𝗿𝗼𝗺 𝗶𝘁𝘀 𝗘𝗕𝗦𝗗 𝗽𝗮𝘁𝘁𝗲𝗿𝗻 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗮 𝗽𝗵𝗮𝘀𝗲 𝗹𝗶𝘀𝘁?
A convolutional neural network from UC San Diego reads crystal symmetry straight from an EBSD Kikuchi pattern, with no candidate phase list. The university reported at least 95% accuracy.
A convolutional neural network from UC San Diego reads crystal symmetry straight from an EBSD Kikuchi pattern, with no candidate phase list. The university reported at least 95% accuracy.