Atomic Loops Research Division
Intelligence from first principles.
Atomic Loops Research is the company’s in-house industrial AI research and development division. It runs two tracks: frontier AI research at TRL 1–3, which investigates open problems in machine learning, and applied AI research at TRL 4–9, which turns those results into production AI systems for manufacturing, logistics and energy operators.
These are the ten subjects under active investigation across both industrial AI research tracks, spanning foundation model alignment, physics-constrained neural surrogates, causal inference, edge inference optimization and federated learning.
Track 01 · TRL 1–3
Frontier AI research is the earlier of the two industrial AI research tracks at Atomic Loops: machine learning problems that have no settled answer yet, at technology readiness levels 1 to 3. Work here is judged on reproducibility and on whether a result transfers to a plant floor, not on a delivery date. Six domains are active, each chosen because an industrial deployment is blocked on it.
Sim-to-real transfer is the problem of making a model trained in simulation hold its accuracy on physical hardware. We study domain randomisation and calibration methods that close the reality gap without a full retrain on plant data.
A physics-constrained model embeds conservation laws directly in its loss function. The result is a surrogate that stays physically valid outside its training distribution, where an unconstrained network drifts.
Causal discovery infers the direction of influence between variables from observational data. Our focus is identifiability under latent confounding, so a maintenance recommendation reflects cause rather than correlation.
Continual learning keeps a deployed model current as the process it monitors changes. The hard part is catastrophic forgetting: retaining last year’s defect classes while absorbing this quarter’s.
A calibrated model reports confidence that matches its observed error rate. We evaluate conformal prediction and deep ensembles against the abstention thresholds a plant operator actually needs.
Neuromorphic systems compute with event-driven spikes instead of dense matrix multiplication. We assess where that trade buys real energy savings on always-on industrial sensing.
Track 02 · TRL 4–9
Applied AI research is the second industrial AI research track: it takes a validated frontier result to technology readiness levels 4 to 9 and runs it against a live process in manufacturing, logistics or energy. Every project carries a measurable production target agreed before work starts, and the four outcomes below are reported against that target rather than against a benchmark.
Surface-defect detection running on the line, measured against operator-adjudicated ground truth rather than a held-out test split.
Time between the first flagged anomaly and the failure it predicted, across rotating equipment under continuous monitoring.
Reduction in mean absolute percentage error against the planning baseline the customer was running before deployment.
End-to-end retrieval and generation latency over technical documentation, measured at the application boundary.
Research protocol
Every industrial AI research project at Atomic Loops moves through the same six stages, in order, on both the frontier and applied tracks. The protocol exists so a result can be reproduced from its manifest by someone who was not in the room when it was produced.
The question is written down as a falsifiable claim with a success metric and a stopping condition, before any modelling starts.
Existing results are reproduced where code is available, so the project starts from a measured baseline rather than a reported one.
Ablations, seeds and held-out splits are fixed in advance. Everything that will be varied is declared before the first run.
Runs are dispatched across the cluster with full configuration capture, so any result can be regenerated from its manifest alone.
Results are re-run under distribution shift, adversarial input and hardware constraints before they are allowed to leave the track.
Findings ship as an artifact package — weights, configs, evaluation harness and known limitations — to production engineering or to the record.
Deployment process
An applied AI research project runs from problem framing to full production in 90 days, with a shadow-mode pilot from roughly day 35. The first two weeks go to framing and a data audit, because most schedule risk in industrial AI comes from data coverage and label quality rather than from modelling. The same five stages apply whether the deployment is a manufacturing line, a logistics network or an energy asset.
Process owners define the decision the model has to improve and the number that proves it improved.
Historian, MES and sensor data are assessed for coverage, label quality and drift before any model is proposed.
Candidate approaches are trained and ablated against the agreed metric, with the simplest sufficient model preferred.
The model runs shadow to the existing process, so its recommendations are scored without acting on the line.
Handover includes monitoring, retraining triggers and an agreed rollback path owned by the customer’s team.
Research outputs
Findings leave the research tracks as an artifact package: weights, configurations, an evaluation harness and a written statement of known limitations. They are released as journal articles, conference papers, technical reports or open-source projects, so a result can be checked by someone outside Atomic Loops.
Embeds conservation laws directly in a surrogate model’s loss function, with the aim of accelerating molecular dynamics simulation.
Takes a weld-defect classifier trained largely in simulation across to physical samples, and examines what the reality gap costs.
Applies structural causal models across supplier tiers so that a forecast reflects propagation through the network rather than correlation.
A toolchain for fitting transformer inference inside a fixed latency budget on ARM edge hardware.
Frequently asked
These five questions cover how the two industrial AI research tracks differ, what a technology readiness level means, how long an applied project takes to reach production, whether external teams can collaborate, and how Atomic Loops publishes its findings.
Frontier AI research investigates open machine learning problems at technology readiness levels 1 to 3 with no delivery date, while applied AI research takes a validated result to levels 4 to 9 against a live production target in manufacturing, logistics or energy. Frontier work is judged on whether a finding reproduces and transfers; applied work is judged on a process metric agreed with the customer before the project starts.
A technology readiness level, or TRL, is a nine-point scale for how far a technology has moved from basic principles to proven operation. TRL 1–3 covers principle, concept and experimental proof; TRL 4–6 covers validation in the lab and then in a relevant environment; TRL 7–9 covers demonstration, qualification and operation in the real system.
Applied AI research projects run on a 90-day envelope from problem framing to full production, with a shadow-mode pilot from roughly day 35. The first two weeks are spent on framing and a data audit, because most schedule risk in industrial AI comes from data coverage and label quality rather than from modelling. The envelope is the same for computer vision on a manufacturing line, predictive maintenance on rotating equipment, and forecasting across a supply chain.
Yes, through an academic track and an industry track. The academic track covers joint publications, thesis supervision and shared compute or data access; the industry track covers applied collaboration against a live production goal, including pilot deployment and full technology transfer to the customer’s own team.
Yes. Findings leave the research tracks as an artifact package containing weights, configurations, an evaluation harness and a written statement of known limitations, and are released as journal articles, conference papers, technical reports or open-source projects. Publishing the limitations alongside the result is what lets an operator judge whether a finding transfers to their own process.
Research partnerships
Atomic Loops runs industrial AI research partnerships on two tracks: an academic track built around a shared publication, and an industry track built around a live production goal in manufacturing, logistics or energy. Both start with a scoping conversation, not a contract.
The academic track is joint frontier research with a university group, structured around a shared publication rather than a deliverable.
The industry track is applied collaboration against a live production goal, with the customer’s process owners in the loop from problem framing onward.
Last updated: