James Le Houx

The implication? Future missions shouldn't just sample random surface ice. We need the instrumentation to pinpoint surface mineralogy and directly probe these natron deposits.

Brilliant work led by @pereralj.bsky.social .
#MaterialsScience #Astrobiology #Synchrotron

If microbial life exists in Enceladus's ocean, natron precipitation could trap those biological signatures, be ejected through the plumes, and deposit near the vent. Natron also offers superior surface preservation against radiation and vacuum sublimation compared to bulk ice. 3/4

We found that natron (Na₂CO₃·10H₂O) precipitates early during fractional crystallisation. As it crystallises, it forms widespread primary fluid inclusions, physically trapping and preserving pockets of the primary ocean fluid within the solid matrix. 2/4

Gemini said A six-panel figure illustrating the real-time freezing of an Na-Cl-CO₃ solution. It features a diagram of the dual-beam experimental setup , a summed diffraction pattern identifying ice, hydrohalite, and natron , and time-lapsed tomography slices alongside corresponding 2D diffraction patterns showing microstructural growth as the temperature drops from 269 K to 205 K. The figure concludes with 3D volume renderings of the final ice, eutectic, and natron phases, plus a line graph tracking the volume evolution of each phase as the temperature decreases.

Enceladus (Saturn's 6th largest moon) vents its subsurface ocean into space. In our latest preprint, we used synchrotron X-ray CT and diffraction to map the rapid freezing of Enceladean brines as they are ejected through tectonic plumes. doi.org/10.21203/rs.... 1/4

The STFC financial situation is worrying. There's a risk we start seeing facilities vs. strategy as a choice.

But you can't build digital twins or AI models if you can't see the physics happening.

Synchrotrons and neutron sources are the ground truth for Net Zero.

#BatterySky #MaterialScience

🌟 Applications are now open for the Next Generation of AI Chemists Programme, a funded opportunity for penultimate-year undergraduates from backgrounds under-represented in chemistry.

📍 Imperial London | 📅 6–7 July 2026 (in person)
👉 Apply here: buff.ly/MLfs9Zg

James Le Houx: Benchmarking Autonomy in Scientific Experiments: A Hierarchical Taxonomy for Autonomous Large-Scale Facilities https://arxiv.org/abs/2601.06978 https://arxiv.org/pdf/2601.06978 https://arxiv.org/html/2601.06978

A chart titled "Cognitive Complexity vs Machine Agency" showing 6 levels of autonomy (0-5). It transitions from Execution (Manual/Scripted) to Perception (Reflexive/Heuristic) and finally to Reasoning (Supervisory/Autonomous). A vertical orange line marks the "Inference Barrier" between Level 2 and 3, and a horizontal dashed line marks the "Liability Threshold" between Level 3 and 4.

Here is how it looks in practice.

The BASE Scale maps Machine Agency against Cognitive Complexity.

Inference Barrier: Where latency dictates if an AI can see in real-time.
Liability Threshold: Where we decide who is liable for accidents.

Benchmarking Autonomy in Scientific Experiments: A Hierarchical Taxonomy for Autonomous Large-Scale Facilities The transition from automated data collection to fully autonomous discovery requires a shared vocabulary to benchmark progress. While the automotive industry relies on the SAE J3016 standard, current ...

I’ve been trying to codify this into a new standard called the BASE Scale.

It’s a 6-level taxonomy (0-5) designed for high-stakes facilities where you only have 48 hours to get results.

I’d love to know if these levels match your experience: arxiv.org/abs/2601.06978

#ScienceSky #AIforScience

We’ve been calling everything from simple scripts to generative agents by the same name.

How do we distinguish between a predictable automation and a stochastic agent that might actually learn something new or drive the stage into a detector?

What's your Liability Threshold?

Autonomous is a catch-all term that means everything and nothing right now.

I'm curious: if you work at a beamline or a shared facility, what is the #1 thing stopping you from letting an AI agent drive your motors?

Is it the latency, the safety, or the lack of trust?

Operando X‐Ray Computed Tomography Reveals the Role of Interfacial Nucleation Nanolayers in Suppressing Mechanical Failure in Zero‐Excess Lithium All‐Solid‐State Batteries This study employs operando X-ray computed tomography combined with an image subtraction method to investigate mechanical failure mechanisms in all-solid-state batteries with lithium and zero-excess ...

Full publication available here: onlinelibrary.wiley.com/doi/10.1002/...

Why do solid-state batteries fail? We used operando X-ray tomography to find out. 🧪

Bare Cu creates localized islands that trigger cracks, but a 50nm Ag nanolayer keeps things uniform and stable, even at high current. 📷🔋

#chemsky #batterysky #sciencesky

By focusing only on mechanistically decisive moments, we can mitigate beam damage and eliminate data redundancy.

Huge kudos to first author Emily Lu and the team across RAL and @diamondlightsource.bsky.social

Indexing for the bots: @arxiv-stat-ml.bsky.social @arxiv-soc-ph.bsky.social

To fix this, we introduce ESME (Entropy-Scaled Measurement Efficiency).

Using Shannon’s information theory, ESME quantifies the rate of information gained per unit of experimental cost.

Essentially, it tells you when to measure, not just where.

A two-panel scientific infographic titled "Conventional vs. Heuristic Operando Experimentation." Panel A shows a conventional pre-programmed approach where periodic measurements (blue circles) fail to capture a battery's failure precursor (red stars). Panel B shows the "Heuristic" approach where an AI Inference Engine, trained on Digital Twins, predicts a failure event and triggers high-rate tomography (green stars) to deterministically capture the failure precursor before a catastrophic event. Bottom text boxes highlight benefits: Efficient Data Acquisition, High-Fidelity Measurements, and Statistically Robust Sampling.

Current battery characterisation faces a crisis of Reliability, Representativeness, and Reproducibility (the 3Rs).

Conventional experiments are passive. We collect terabytes of data but often miss the transient events that actually kill cells like the split-second of dendrite nucleation.

Autonomous battery research: Principles of heuristic operando experimentation Unravelling the complex processes governing battery degradation is critical to the energy transition, yet the efficacy of operando characterisation is severely constrained by a lack of Reliability, Re...

We need to stop measuring experimental success by data volume (TB/hour) and start measuring it by scientific insight.

In our new arXiv preprint, we propose Heuristic Operando experimentation: a framework to teach beamlines how to think.

arxiv.org/abs/2601.00851 #ScienceSky #AI #BatteryResearch

Some news! I’m starting a new chapter as a Senior Lecturer at Greenwich and launching the BASE Lab.

We are teaching beamlines how to think, building a Self-Driving Microscope to hunt for battery failure.

I'm also hiring our founding PhD student. Come build with us! #SciSky #AIforScience

​In the wake of the Airbus A320 groundings, it's worth remembering that the ISIS Neutron and Muon Source has a facility dedicated to replicating this kind of cosmic bombardment.

​It’s called ChipIR.

​It allows engineers to test the degradation of avionics before they ever leave the ground.

We need to stop looking backward at incomplete datasets and start building semantic standards for the future.

That is the focus of our new initiative, the Faraday Engine. Mission One is Health; we are working hard to ensure Energy is ready to follow in 2026. 🔋

We have petabytes of legacy data, but it is mostly "What" (characterisation data). To build predictive battery models, we need the "How" (synthesis, thermal history).

Without that semantic link, GenAI cannot learn the recipe. It sees the destination but doesn't know the route. 2/3

Diving into the new UK AI for Science Strategy.

They’ve made a definitive call, prioritise generating new data over optimising existing archives.

As a materials scientist, I think this is absolutely the right call. Here is why. 🧵1/3

#AIforScience #MaterialsScience

Played around with Meta's new SAM 3 on our X-ray data today.

Top: Our manually trained U-Net (accurate but slow). Bottom: SAM 3 out-of-the-box ("white objects").

It misses some outliers, but getting 90% of the way there with zero training is a win for high-throughput workflows.

#Science #AI

Huge congrats to Harriet Jones for winning best poster at #IBSim-4i!

She's leading our project adapting #code_saturne for on-the-fly simulation from CT data enabling autonomous workflows.

Thx judges & sponsor #3Dmagination!

#IBSim4i #Simulation #Physics #CFD #AutonomousScience #chemsky #BattChat

At PorTo, ILL's new beamline, tackling Li-ion failure: electrolyte consumption, interlayer buckling, & overcharge.

Running high-res neutron imaging + multi-modal (X-ray/Neutron) correlation & unsupervised AI.

Thx Bratislav Lukić!

#Liion #ILL #CorrelativeImaging #DataScience #chemsky #BattChat

ACS Spring 2026 abstract deadline: September 29th!

We're chairing a new session in Atlanta. If you use large-scale facilities to study energy storage materials with photons, neutrons, muons, or lasers, we want to hear from you.

Submit here: tinyurl.com/4e3envar

#chemsky #research #ACSSpring2026

EIS ice baby

New ideas can be found in the intersection between battery science and planetary science.

Myself and @thebeamline.bsky.social are exploring how ice microstructure will impact radar measurements on Europa using tomography and EIS.

#Europa #icymoons #planetaryscience #tomography

🧪

In-situ dark field X-ray microscopy of solid state batteries this week. Seriously challenging experiment, thank you to the whole team at ID03 @esrf.fr

This was a huge team effort, so massive thanks to my co-authors and the teams at @psich.bsky.social, ISIS Neutron and Muon Source, The Faraday Institution, and Diamond Light Source. Go read the paper!