Master-Thesis-2024

Local thermal-impulse campaign and V8.3 closed-loop follow-up

Thermal-impulse dates: 22–24 June 2026
Closed-loop run: 25–26 June 2026
Scope: local thermal sensitivity, followed by a phase-resolved forensic assessment of V8.3 feedback behaviour.

Read this report as evidence, not a success claim. The thermal-impulse campaign identifies likely disturbance pathways. V8.3 demonstrates an initially bounded feedback interval, but it does not demonstrate robust long-duration centroid stabilisation.

---

1. Thermal-impulse result: what is sensitive?

Each accessible location received the same nominal thermal input:

P = 2.20 W, Δt = 70 s, E = 154 J

Location	Peak absolute ΔY	Peak radial image response	Peak absolute ΔOPL	Interpretation
Grating A (G2)	0.888 px	1.047 px	0.142 µm	Largest detector-plane response
Grating B	0.200 px	0.283 px	0.054 µm	Smaller, clean comparison location
Camera mount A / right mount	0.713 px	0.941 px	2.273 µm	Largest OPL response
Camera mount B	0.634 px	0.652 px	0.116 µm	Fast image response; strongest rigid-shift validation

Working hardware interpretation: Grating A is the priority image-motion sensitivity location; Camera mount A is the priority OPL-sensitivity location. These are different observables and do not establish a single hardware culprit.

Independent FITS check

Location	FITS frames	Full-record MaxIm–FITS discrepancy RMS	Interpretation
Grating A	41	0.183 px	Trend supported; amplitude remains method-dependent
Grating B	72	0.098 px	Strong image-motion validation
Camera mount A	72	0.389 px	Strong OPL result; later centroid behaviour needs caution
Camera mount B	36	0.075 px	Strongest rigid-shift validation

Caution: environmental telemetry gives context, not local component temperature. It cannot by itself identify detector-mount, grating-substrate, or camera-cooling gradients.

---

2. Why V8.3 follows the thermal-impulse campaign

The two experiments answer complementary questions:

Thermal-impulse campaign	V8.3 feedback run
Which accessible location creates image or OPL sensitivity?	Can the controller maintain image stability while the instrument evolves naturally?
Local pulse and passive relaxation	Long-duration closed-loop operation
Identifies disturbance pathways	Tests controller authority, model validity, and actuator response

The V8.3 analysis therefore uses the impulse campaign to motivate local-temperature sensing and model re-identification, not to claim that a single global TEC setpoint can compensate every thermal-optical pathway.

---

3. V8.3 exact timeline

Times below are timestamps recorded by the control PC; the CSV did not retain an explicit timezone.

Stage	Start	End	Duration
Passive warm-up	25 Jun 2026, 22:51:20	25 Jun 2026, 23:54:37	63.3 min
TEC identification, positive leg	25 Jun 2026, 23:55:47	26 Jun 2026, 00:13:52	18.1 min
TEC identification, return leg	26 Jun 2026, 00:15:03	26 Jun 2026, 00:21:30	6.5 min
Post-identification settling	26 Jun 2026, 00:22:41	26 Jun 2026, 00:36:53	14.2 min
TEC-only control	26 Jun 2026, 00:38:04	26 Jun 2026, 01:06:29	28.4 min
MIMO phase	26 Jun 2026, 01:07:46	26 Jun 2026, 11:45:56	10.64 h

Figure 1. Full V8.3 telemetry. The horizontal axis is elapsed experiment time. Warm-up is passive logging; TEC-only permits thermal setpoint correction only; MIMO enables the outer supervisory allocation logic. The later divergence must not be hidden by the early stable segment.

---

4. What the run actually achieved

Segment	Frames	Duration	Radial RMS	P95 radial error	Within ±0.5 px	TEC actions	AO actions
TEC-only	23	28.4 min	0.092 px	0.149 px	100.0%	5	0
MIMO, feedback clock ≤ 6 h	255	5.47 h	0.225 px	0.298 px	99.6%	29	6
MIMO, feedback clock > 6 h	240	5.15 h	1.440 px	2.286 px	14.6%	10	0
Full MIMO	495	10.64 h	1.016 px	1.977 px	58.4%	39	6

The early MIMO interval is a real bounded interval. It is not, by itself, causal proof that outer feedback produced the stability: V8.3 did not include a matched no-outer-control baseline under comparable conditions.

The hardware TEC loop itself continued to track its own target: median absolute measured-target gap was 2.79 mK in early MIMO and 3.52 mK in late MIMO. The late centroid excursion was therefore not evidence that the inner TEC temperature regulator had simply stopped tracking.

4.1 Histogram view: whole run versus stable first six feedback hours

Figure 2. Histogram comparison of centroid and radial-error distributions for the full V8.3 run versus the stable first six feedback hours. The stable early interval is tightly concentrated, whereas the full-run distribution broadens substantially because of the later drift.

4.2 Distribution shift: early stability versus late drift

Figure 3. ΔY distribution shift between the first six feedback hours and the later drift interval. The early distribution is narrow and centred close to zero, whereas the late interval shifts strongly negative, consistent with the persistent centroid excursion.

4.3 Whole-run histogram summary

Figure 4. Whole-run histograms of ΔX, ΔY, and radial error. Median and mean markers show that ΔY and radial error are strongly affected by the late-drift tail, even though a substantial fraction of frames remain near the early stable operating region.

---

5. Forensic answer: why did long-duration stability fail?

5.1 The onset was a control-model mismatch, not a proven external environmental step

The first persistent loss of the 0.5 px containment band began at 26 Jun 2026, 07:05:38. At 07:30:12, the filtered image state changed from the preceding frame by:

Quantity	Change over 78 s
ΔX	−0.174 px
ΔY	−0.333 px
Radial error	+0.372 px
OPL residual	−0.011 µm
ECU temperature	0.000 °C
ECU pressure	−0.05 hPa
ECU water content	−0.147 g m⁻³
BME temperature	−0.16 °C

No logged temperature, pressure, humidity, or OPL discontinuity matches the centroid jump in magnitude or timing. An unmeasured local thermal-mechanical or centroid-measurement disturbance remains possible, but the logged data do not identify an environmental channel as the unique trigger.

Figure 5. Around the late-drift onset, the uncapped PI-D request became non-zero and large, but many proposed commands were vetoed by the predicted-cost gate. OPL and pressure evolved continuously; they do not show an equivalent abrupt event at the 07:30 centroid transition.

5.2 The implemented model had insufficient verified authority

The active thermal column remained fixed throughout MIMO at approximately:

B_T = [0, +7.073, 0]^T per °C

In practical terms, the controller model assumed that TEC could directly correct Y only, with no identified direct authority in X or OPL. A single TEC setpoint is therefore not able to independently drive ΔX, ΔY, and ΔOPL to zero. The system was structurally rank-limited until a validated AO response was available.

At 07:30:12, the outer PI-D terms requested approximately +25.9 mK before limits. The provisional-model policy limited a single action to +4 mK. The predicted MIMO cost would then improve only from 222.27 to 210.55, i.e. 5.3%, below the required 10% reduction. The command was therefore vetoed as no_predicted_mimo_gain.

At 07:32:47, radial error exceeded 1 px and the integral state entered coarse_error_no_integrate. This was intentional anti-windup, but it removed the mechanism that could accumulate a persistent correction. The result was a conservative recovery deadlock: the controller calculated a non-zero request but refused the bounded command because its one-step predicted improvement was too small.

5.3 The thermal model later failed its own response validation

At 08:42:34, a +4 mK TEC step was followed by a measured +14.3 mK thermal movement. The centroid response was opposite to the fixed model prediction and was logged as direction_mismatch. The correct interpretation is not that the hardware necessarily failed; it is that the local thermal response used by the controller was no longer predictive at that operating point.

5.4 AO did not provide an independent recovery path

Figure 6. All six AO moves were +X steps. The next-frame observed centroid changes were much smaller than, and often directionally inconsistent with, the stored AO calibration. Once cumulative X travel reached +6 steps, the software’s global soft-travel condition prevented further AO use on either axis.

AO finding	Evidence
Y was never selected	All six commanded moves were (+1, 0) in the software step convention
Stored AO calibration was not validated by the run	Observed next-frame responses were far below the predicted displacement and did not consistently align in direction
X travel disabled Y	The policy used a global maximum travel test; reaching +6 in X blocked both X and Y
AO could not rescue late drift	AO was restricted to fine residuals and was unavailable once error left that region

---

6. Defensible conclusion from V8.3

V8.3 showed an early bounded interval but did not provide robust long-duration centroid stabilisation. The late failure is best explained by a changed or unmodelled plant response combined with restrictive supervisory-control gates and an unverified AO calibration. The available telemetry does not identify a unique logged environmental trigger.

This is more precise than either of the following unsupported claims:

• “the feedback worked for the whole run”
• “the problem was only environmental”

---

7. Version 9 requirements derived from the evidence

1. Bidirectional thermal identification: do not arm autonomous feedback from a one-leg or inadequately settled temperature response.
2. Change-detection mode: a direction_mismatch must freeze ordinary correction and trigger bounded re-identification, not normal planning against an invalid model.
3. Reachability check: explicitly test the rank and conditioning of the TEC–AO response matrix before claiming three-output MIMO regulation.
4. AO recalibration in the current run: identify X and Y separately; do not consume travel after a failed validation; use independent axis travel limits.
5. Recovery logic without deadlock: compare the predicted benefit of a step against an absolute noise-aware threshold, not only a fixed relative cost reduction when the residual is already large.
6. Matched evidence run: retain a fixed-reference passive segment before controlled operation and report the same metrics for both.
7. True resume checkpoint: save reference, model coefficients and covariance, filter buffers, integrator state, pending actuator validation, action ledger, and AO state; resume only after a passive continuity check.
8. Local disturbance measurement: prioritise temperature sensing at the Camera mount A/right-mount and Grating A sensitivity regions indicated by the thermal-impulse campaign.

---

Public boundary

This public report contains derived metrics and figures only. It omits raw telemetry, detailed hardware geometry, optical alignment, operational communications, and live controller implementation details.