Forecasting M-class flares from the EUV image.

Three image branches and a magnetic foundation

Helios reads four streams. Three are EUV imagery from SDO/AIA: the 131 Å channel (sensitive to ~10 MK flare plasma), the 193 Å channel (corona, ~1.5 MK), and the 304 Å channel (chromosphere/transition region, ~80 000 K). The fourth is HMI line-of-sight magnetograms at ~720 s cadence. The combination matters: the EUV channels carry the structural signature of pre-flare reorganisation; the magnetogram carries the photospheric driver. Magnetograms alone get to roughly TSS 0.65 at 30 minutes on our holdout; adding the AIA branches in 2026-04 added +0.06 TSS at the 30-minute horizon, which is the headline improvement of the current release.

A spatiotemporal CNN with a ConvLSTM head

The backbone is a four-branch convolutional encoder (one branch per AIA channel plus one for HMI). Each branch processes a 256×256 patch centred on the active region at 720-second cadence over a 6-hour input window — that is 30 frames per channel, 120 frames in total per active region per forecast. Branch outputs concatenate at a 32×32 spatial resolution and feed into a two-layer ConvLSTM head that produces the temporal forecast. A small fully-connected layer reads the ConvLSTM final state and emits the probability of a ≥M-class flare in each of four forecast bands: 0–30 min, 30 min–2 h, 2 h–6 h, 6 h–24 h.

The probability is not the raw network output. Raw softmax probabilities from this kind of network are systematically over-confident, and an operations desk that acts on them ends up with a calibration crisis after the first quiet week. We post-process with isotonic regression fitted on a held-out calibration year (2024); the calibration is re-fitted with every release.

What the model has seen

The validation/calibration/holdout split is by calendar year and not by active region; we are aware that this is the easier split. We also report a leave-one-region-out evaluation for the holdout year on the validation notebook, where TSS at 30 minutes drops to 0.66 — that number is the one we will defend in a contested room.

What Helios is not good at

Limb events. Active regions within ~15° of the solar limb are systematically harder for Helios because the line-of-sight projection of the magnetogram becomes unreliable and the AIA branches see foreshortened structure. We flag these forecasts internally with a confidence-reducing multiplier; the validation notebook reports limb performance separately.

Multi-flare days. When two strong active regions are co-evolving on the disk and one of them produces a precursor that propagates structurally to the other, the network attributes the precursor to the wrong region. The failure shows up as a "near miss" in the per-region confusion matrix — the flare did happen, just not where Helios said.

Sustained quiet conditions. Calibration drifts after long quiet periods; we trigger a re-calibration when the rolling-30-day base rate falls below a threshold. The trigger has fired three times in the holdout year.

Where Helios runs today

One pilot deployment in shadow-traffic mode at the ESA Space Situational Awareness Service Coordination Centre, comparing Helios outputs against the operational baseline before any operational decision is allowed to be influenced. Shadow traffic began 2026-02 and is currently scheduled to run through 2026-Q4 with quarterly review. Two further pilots — at a regional space-weather centre and at an autonomous-observatory scheduling team — are in negotiation.

Helios is delivered as a Docker container with a deterministic CUDA build (pinned to NVIDIA NGC 24.07-cuda12.5-pytorch2.3). Inference latency on an A100 is 380 ms per active region per forecast cycle; on an L4 it is 620 ms. The validation notebook reproduces the headline TSS on a single GPU in under thirty minutes.