SCS-Lab/FUSE-PED

FUSE-EFIT (in era), ONNX

A 20 Hz, strictly causal model of the DIII-D plasma edge, driven by the heating, fuelling and field actuators, the measured plasma current, the loop voltage and the EFIT01 boundary shape. Physical units in, physical units out. Code, documentation and the input contract: https://github.com/NoahH72/FUSE-Pedestal-Predictor-Mamba-ONNX (docs/EXPORT.md, docs/INPUT_CONTRACT.md, MODEL_CARD.md).

Two forms, one calling convention

form	what it is	networks per tick	ONNX files	step call, 1 thread	step call, 8 threads
single	replicate 0 of the in-era finals (seed 100), the frozen delivered network	1	184 MB	4.87 ms (99th percentile 5.08 ms)	1.95 ms (99th percentile 2.00 ms)
averaged	the eight in-era finals (seeds 100-107); the mean of their raw outputs at every tick	8	1468 MB	56.33 ms (99th percentile 59.45 ms)	18.25 ms (99th percentile 18.60 ms)

Timing: Intel(R) Xeon(R) CPU Max 9470C, onnxruntime 1.30.0, CPU, batch 1. The control cycle is 50 ms per tick. The two forms take and return the same tensors with the same names, shapes and units; only the recurrent state is larger for averaged (19.9 MB per shot against 2.5 MB). A host switches form by choosing a folder.

Files

era/single/fuse_efit_single_step.onnx        one 50 ms tick per call, state in and out
era/single/fuse_efit_single_sequence.onnx    one window of 256 ticks
era/averaged/fuse_efit_averaged_step.onnx
era/averaged/fuse_efit_averaged_sequence.onnx
era/<form>/model_config.json                 the input and output contract, machine-readable
era/<form>/norms.json                        the statistics baked into the graphs
era/<form>/input_ranges.json                 training ranges per channel, for a units check
era/<form>/provenance.json                   source checkpoints with sha256, code identity, versions
SHA256SUMS                                   sha256sum -c SHA256SUMS

Download and check

git clone https://github.com/NoahH72/FUSE-Pedestal-Predictor-Mamba-ONNX
cd FUSE-Pedestal-Predictor-Mamba-ONNX
pip install numpy onnxruntime huggingface_hub
hf download SCS-Lab/FUSE-PED --local-dir export        # this page's files, into the code's export/ folder
cd export && sha256sum --check SHA256SUMS && cd ..
python examples/verify_integration.py --bundle export/era

Every line of the checksum check must say OK, and the integration check must pass, before the model is used.

Run it

from fuse_efit.export.runtime import FuseEfitPredictor     # numpy + onnxruntime only

model = FuseEfitPredictor("export/era", form="single")       # or form="averaged"
state = model.init_state()                                   # zeros at the start of every shot
for each 50 ms tick:
    y, state = model.step(inputs_25, indicators_5, state)    # exactly once per tick

Step graph tensors: inputs (batch, 25), indicators (batch, 5), available (batch, 25), conv_state, ssm_state in; predictions (batch, 9), conv_state_out, ssm_state_out out.

Inputs (physical units, this order)

#	channel	unit
0	pinj	MW
1	tinj	N m
2	ech_mw	MW
3	gasa_cal	Torr L/s
4	gasb_cal	Torr L/s
5	gasc_cal	Torr L/s
6	gasd_cal	Torr L/s
7	gase_cal	Torr L/s
8	bt	T
9	ip	MA
10	aminor	m
11	rsurf	m
12	zsurf	m
13	kappa	dimensionless
14	tritop	dimensionless
15	tribot	dimensionless
16	drsep	m
17	rvsout	m
18	gapin	m
19	rvsou	m
20	rvsiu	m
21	shafranov	m
22	zeta_ou	dimensionless
23	zeta_ol	dimensionless
24	vloop	V

Indicators, 0/1 per tick: efit_valid, vloop_valid, rvsout_present, rvsou_present, rvsiu_present. Availability, 0/1 per shot and channel (default 1). Where an indicator or the availability flag is 0 the value is ignored; elsewhere supply a real value (not-a-number is read as the training mean).

Outputs

#	head	unit
0	ne	1e19 m^-3
1	te_ped	keV
2	ti_ped	keV
3	t_rot_ped	krad/s
4	ne_knee	1e19 m^-3
5	te_knee	keV
6	rho_sym	rho (normalised radius)
7	ne_top_loc	rho (normalised radius)
8	hmode	probability in [0, 1]

ne_knee, te_knee, rho_sym and ne_top_loc are defined only in H-mode: report them only where hmode is at or above a threshold. Each form has its own, set on validation: single 0.54 (balanced) and 0.5 (conservative); averaged 0.53 and 0.47. On the test shots each keeps about 95-96 % of H-mode ticks with about 5 % L-mode among those kept (MODEL_CARD.md §6).

Accuracy

Typical error per test shot (the median over in-era test shots of each shot's RMSE, physical units; hmode: its loss), from the single opening of the test set:

output	unit	single	averaged
ne	1e19 m^-3	0.2928	0.2594
te_ped	keV	0.05405	0.04808
ti_ped	keV	0.09957	0.08967
t_rot_ped	krad/s	4.043	3.678
ne_knee	1e19 m^-3	0.3279	0.2973
te_knee	keV	0.06058	0.05636
rho_sym	rho (normalised radius)	0.01047	0.009892
ne_top_loc	rho (normalised radius)	0.0125	0.01188
hmode	loss	0.1373	0.1251
combined loss		0.9889	0.8986

Export accuracy

Every graph agrees with the PyTorch model to 7.4e-06 of each head's training standard deviation or better, on 1,988 validation ticks per form and 2,000-tick runs.

Limits

Needs EFIT01 geometry, the between-shot reconstruction, as trained; it is not for use during the shot. Fed real-time EFIT geometry instead it is about 30 % worse, and far worse on shots where the real-time reconstruction lacks the strike points. No uncertainty estimate. A unit error gives confident nonsense, not an error: run examples/verify_integration.py and check_units once. Full list: MODEL_CARD.md §8.

Licence and contact

Licence: MIT (the LICENSE file). Contact: Noah Harris, harrnoah@oregonstate.edu. Code: https://github.com/NoahH72/FUSE-Pedestal-Predictor-Mamba-ONNX.