111047 euv-lithography-fault-detection-classification semiconductor engineering

**Fault Detection and Classification for EUV Lithography** # Fault Detection and Classification for EUV Lithography ## Introduction Fault Detection and Classification for EUV Lithography is an engineering workflow for high-resolution pattern transfer. Its purpose is to detect abnormal operation and assign actionable fault classes. A useful implementation joins process knowledge, trustworthy measurements, statistical validation, and explicit decision rules; a model score alone is not an operational result. The primary evidence includes dose, focus, mask data, resist chemistry, and wafer inspection maps. Each source needs an owner, unit, timestamp policy, calibration state, valid range, and product or equipment context. The principal performance measure is **detection recall and false alarms per lot**. The main failure mode to guard against is **novel faults that do not match trained classes**. ## Problem Definition Define the decision before selecting an algorithm. Record who acts, when the decision is made, what alternatives are allowed, and the costs of false positive, false negative, and delayed action. Separate controllable recipe inputs from observed states, outcomes, and contextual variables such as product, chamber, route, and maintenance age. Let $x_t$ be the measured state, $u_t$ the controllable setting, $y_t$ the outcome, and $c_t$ the manufacturing context. A basic predictive formulation is $$ \hat y_t=f_\theta(x_t,u_t,c_t), \qquad r_t=y_t-\hat y_t. $$ For decision support, minimize expected loss subject to the qualified operating envelope: $$ u_t^*=\arg\min_{u\in\mathcal U}\;\mathbb E[L(y,u)\mid x_t,c_t] \quad\text{subject to}\quad g_j(x_t,u)\leq 0. $$ Constraints represent safety, process integration, equipment, and product rules. They should remain enforceable if the analytical service is unavailable. ## Data and Measurement Strategy Create a versioned data contract for every signal. Check units, clocks, sampling rate, missingness meaning, censoring, detection limits, and joins between wafer, lot, tool, chamber, recipe, and metrology identifiers. Preserve raw values and record transformations rather than overwriting questionable observations. Use chronological splits and keep lots, wafers, or dies from the same physical group in one split. Random row splits often leak spatial and temporal information. Compare the proposed method with the current operating rule, a last-value baseline, and a transparent statistical model. Recommended data-quality gates include: - timestamp and genealogy consistency; - calibration and maintenance-state validity; - physically plausible ranges and rates of change; - missing-channel and stale-signal detection; - product, tool, and operating-regime coverage; - immutable lineage from source to deployed feature. ## Modeling Approach Start with interpretable control charts, generalized linear models, trees, or state-space models. Add nonlinear, deep, or hybrid models only when validation shows material benefit. Encode known symmetries, monotonic relationships, conservation rules, and feasibility constraints where appropriate. Quantify uncertainty using bootstrap ensembles, Bayesian inference, conformal prediction, or calibrated quantile models. Evaluate both accuracy and calibration: $$ \mathrm{RMSE}=\sqrt{\frac1n\sum_i(y_i-\hat y_i)^2},\qquad \mathrm{Coverage}=\frac1n\sum_i\mathbf 1\{y_i\in[\ell_i,u_i]\}. $$ If interventions are proposed, prediction is insufficient. Use designed experiments or a defensible causal design to estimate what changes after an action. Document assumptions and negative controls. ## Implementation Workflow 1. Frame one bounded decision and define its owner, cadence, baseline, and acceptance threshold. 2. Build validated feature views from the governed manufacturing record. 3. Train a simple baseline and then candidate models using time-aware evaluation. 4. Stress-test missing signals, tool changes, product changes, maintenance events, and rare extremes. 5. Run in shadow mode and capture recommendations, operator responses, latency, and eventual outcomes. 6. Introduce bounded authority with approval gates, rate limits, feasibility checks, and rollback. 7. Monitor data, predictions, actions, and delayed outcomes as one closed loop. Every release should pin code, training data, feature definitions, environment, random seeds, and decision policy. Store the previous deployable artifact and rehearse rollback. ## Evaluation and Acceptance Report detection recall and false alarms per lot by time period, product, tool, chamber, recipe family, and relevant spatial region. Include confidence intervals and the number of independent lots, not only the number of rows. Test tail behavior because average accuracy can conceal costly excursions. An acceptance package should cover: - improvement over operational and statistical baselines; - calibration of confidence or prediction intervals; - stability across seeds and adjacent hyperparameters; - inference latency and resource use on target infrastructure; - abstention behavior for out-of-distribution inputs; - recovery during network, sensor, and service failures; - review and sign-off by process, equipment, quality, and manufacturing owners. ## Deployment Architecture Keep acquisition, validation, feature computation, inference, policy, and actuation as separately observable stages. The fast safety path must not depend on a cloud model. Publish analytical recommendations through versioned schemas with explicit units, timestamps, confidence semantics, expiry times, and idempotent retry behavior. Begin with offline replay, then shadow operation, then a limited canary on representative equipment. Expand only after stable evidence. Log the complete decision context so an engineer can reconstruct why a recommendation was issued. ## Monitoring and Failure Handling Monitor input drift, missingness, residuals, calibration, action frequency, overrides, process outcomes, and detection recall and false alarms per lot. Segment alerts by product and equipment context. Define warning, abstain, and shutdown thresholds before deployment. The risk of novel faults that do not match trained classes deserves a dedicated stress test and response playbook. When inputs are invalid or outside validated support, the system should abstain, preserve evidence, notify the accountable owner, and fall back to the qualified baseline. Never silently substitute a convenient value for a safety-relevant measurement. ## Practical Example Select one tool group and one product family with reliable genealogy. Assemble a forward-chaining development period, a later validation period, and an untouched qualification period. Train the baseline and candidate model, then replay both against historical decisions. During shadow mode, compare recommendations with actual engineering disposition and downstream results. A successful pilot demonstrates repeatable improvement in detection recall and false alarms per lot, calibrated uncertainty, acceptable review load, and safe degradation. If improvement disappears after controlling for time, product mix, or maintenance state, treat that result as evidence of confounding rather than tuning the test until it passes. ## Key Takeaways - Fault Detection and Classification for EUV Lithography should begin with a governed manufacturing decision, not a preferred model. - For EUV Lithography, trustworthy context and genealogy are as important as algorithm choice. - Validate chronologically and by independent physical groups. - Pair point predictions with calibrated uncertainty and explicit abstention. - Deploy gradually with bounded authority, monitoring, and a tested fallback. - Optimize detection recall and false alarms per lot while actively testing for novel faults that do not match trained classes.