Closing the Scale Gap: What AI Must Carry from Process Data to Lifetime Decisions

Photovoltaics has no shortage of data, far from this. The scale gap appears when evidence created for one purpose is asked to support a different decision.

A process model identify a drift. An image classifier detect an anomaly. A degradation model fit a convincing trajectory. Yet each result lose its industrial meaning when it moves from a tool to a material behavior, from a signal to a failure mechanism, or from an accelerated test to a lifetime assumption.

This is the central difficulty in applying artificial intelligence across photovoltaic industrialization: prediction is not a direct evidence transfer.

Deployment-grade AI must preserve the identity, physical meaning, uncertainty and decision context of the evidence it uses. Without that continuity, even the most strong model can still support the wrong action.

The scale gap is a loss of data meaning

A data point does not have the same meaning at every scale. Chamber pressure, source power or substrate temperature are measurements of equipment and at the same time represent process conditions; they are not direct measurements of the material characteristics and state created by the process. An electroluminescence image records a radiative response under specified acquisition conditions; it is not, by itself, a diagnosis of a potential degradation root cause or future progression. An accelerated test records behaviour under a defined and concrete stress history; it does not automatically establish a service-life distribution under real conditions climate wise

AI can help connect these domains. But the connection is valid only if the evidence required at the next scale is carried forward rather than replaced by a model score.

1. Process data to material and product state

A manufacturing line can remain inside its nominal recipe window while the produced materials characteristics drift. Tool condition, chamber history, incoming-material variation, maintenance, calibration and interactions between parameters may change the output without a single setpoint crossing an alarm limit.

A useful model will may be detect that multivariate drift earlier than conventional limits. However, the industrial question is not simply whether the anomaly score moved, but whether the score can be linked to a physically meaningful state changing product characteristics and more importantly to a controlled action.

That requires product and lot identity, recipe and software version, tool and maintenance history, time alignment, metrology calibration, sampled ground truth and downstream performance. It also requires a defined decision: increase sampling, hold a lot, adjust a process window, condition a tool or release production.

Without those links, AI can reveal correlation while leaving the causal object, and most imprtantly the accountable decision, uncertain.

2. Signal to mechanism

Electroluminescence classification provides a concrete example. Deitsch and co-authors demonstrated automated classification of defective PV cells using high-resolution EL images. The work established that image-based methods can support scalable inspection. It did not claim that a visual class, on its own, proves a unique degradation mechanism or a lifetime outcome.

This distinction matters because similar image signatures can arise from different physical causes, while some visible abnormalities may have limited electrical or progression significance. Classification answers: what does this observation resemble within the labelled population? Mechanism requires additional evidence: why did it occur, how does it affect the device, and under which stresses will it progress?

The bridge normally requires more than one modality: acquisition conditions, I–V response, production cohort, bill of materials, process history, targeted stress testing, field evolution and, where justified, selective physical analysis. The purpose is not to turn every inspection into a research programme. It is to distinguish screening, triage and root-cause decisions—and to avoid converting a convenient label into an unsupported lifetime claim.

3. Test response to lifetime decision

The consensus statement on ISOS-based stability assessment for perovskite photovoltaics illustrates why procedural consistency and reporting matter. The authors aimed to improve comparability, reproducibility and understanding of failure modes; they explicitly did not present the protocols as a replacement for qualification standards.

The 2025 IEA PVPS Task 13 report on degradation and failure modes in new cell and module technologies reinforces another point: degradation behaviour depends on technology, materials, architecture, stress combination and operating conditions. Laboratory-to-field replication remains a technical problem, not a data-format problem.

A fitted degradation curve becomes decision-relevant only when the tested configuration is traceable, the stress dosage and operating point are known, the dominant mechanism is sufficiently understood, the acceleration rationale is defensible, uncertainty is visible and field evidence can challenge the assumption. Otherwise, model precision can hide extrapolation risk.

FAIR is necessary—but it is not physical validation

The FAIR Guiding Principles provide an essential foundation for making data findable, accessible, interoperable and reusable, with particular attention to machine actionability and provenance.

For industrial AI, those properties are necessary. They allow a model output to be traced, combined and revisited. But FAIR data can still be unrepresentative, poorly calibrated, incorrectly labelled or physically insufficient for the intended decision. FAIR does not establish that a correlation is causal, that a test reproduces field exposure, or that a model remains valid after a product, factory or environment changes.

The practical requirement is therefore FAIR plus physics, uncertainty and decision ownership.

A decision–evidence matrix for cross-scale AI

The matrix below is not a universal acceptance standard. It is a way to expose what must survive each transfer before an AI result is used to change an industrial decision.

Scale transfer	Evidence that must travel with the output	Decision protected
Process → state	Product and lot identity; tool and recipe version; time alignment; calibration; sampled ground truth; downstream correlation	Adjust, maintain, hold or release
Signal → mechanism	Acquisition conditions; electrical impact; BOM and production cohort; stress or field progression; targeted confirmation	Screen, triage, investigate or change a control
Test → lifetime	Tested configuration; stress dosage; operating point; mechanism; acceleration rationale; uncertainty; field anchors	Revise qualification, degradation assumptions or monitoring

Six questions before an AI result enters an industrial workflow

• What exact decision will change if the output crosses its threshold?

• Is the decision population represented by the training and validation data—technology, factory, product configuration and operating environment included?

• Can every prediction be traced to a sample, product, process history, measurement condition and model version?

• What physical hypothesis makes the relationship plausible, and what evidence distinguishes competing mechanisms?

• How are measurement uncertainty, model uncertainty and out-of-distribution inputs made visible to the decision owner?

• What new evidence would invalidate the model, trigger escalation or require requalification?

If these questions cannot be answered, the system may still be a useful research or diagnostic tool. It is not yet a deployment-grade decision system.

The industrial implication

AI does not close the PV scale gap merely by increasing prediction accuracy. It closes part of the gap when it allows evidence to remain meaningful as it moves across teams, equipment, products, tests and operating years.

That places AI inside process control, change management, qualification and reliability workflows—not beside them as an isolated digital pilot. The model must be connected to a decision owner, a validation boundary and a feedback path. When a recipe, material, module architecture, factory or environment changes, the evidence chain must show what remains transferable and what must be demonstrated again.

This is the question I am developing for my forthcoming contribution to WCPEC-9 in Daejeon: not how to add more AI to photovoltaics, but what AI must preserve when evidence moves from process data to lifetime decisions.

The scale gap is not only a shortage of data. It is what happens when data crosses a boundary and loses the meaning needed for a responsible decision.

Selected references

• Deitsch, S. et al. (2019). Automatic Classification of Defective Photovoltaic Module Cells in Electroluminescence Images. Solar Energy 185, 455–468. https://doi.org/10.1016/j.solener.2019.02.067

• Wilkinson, M. D. et al. (2016). The FAIR Guiding Principles for scientific data management and stewardship. Scientific Data 3, 160018. https://doi.org/10.1038/sdata.2016.18

• Khenkin, M. V. et al. (2020). Consensus statement for stability assessment and reporting for perovskite photovoltaics based on ISOS procedures. Nature Energy 5, 35–49. https://doi.org/10.1038/s41560-019-0529-5

• IEA PVPS Task 13 (2025). Degradation and Failure Modes in New Photovoltaic Cell and Module Technologies. https://iea-pvps.org/key-topics/degradation-failure-modes-new-cell-module-technologies/