NeSySpec: Neuro-Symbolic Technical Intelligence for Auditing Engineering Specifications with an Application to Aircraft

Pier Mattia Basciano, Gianvito Antenori, Cristiano Andolfi, Giovanni Farina, Salvatore MonteleonearXiv 2026 · arXiv

Engineering specifications are the DNA of complex systems like aircraft. But what happens when that DNA is mutated—whether by simple data entry errors, incomplete records, or deliberate falsification? For engineers auditing these specifications, the task is a needle-in-a-haystack problem: they must identify which physical constraints are violated, pinpoint the erroneous data field, and determine how to fix it, all while verifying the data against visual evidence like photographs or satellite imagery.

A new paper titled "NeSySpec: Neuro-Symbolic Technical Intelligence for Auditing Engineering Specifications" introduces a framework designed to automate this entire audit pipeline. By blending the reasoning power of logic with the pattern-recognition capabilities of deep learning, the authors offer a tool that doesn't just flag errors but explains them.

TL;DR

NeSySpec is an integrated neuro-symbolic framework built on Logic Tensor Networks that audits engineering specifications by checking them against learned physical constraints and multimodal evidence. It outperforms classical statistical methods and large language models in identifying violated constraints, localizing erroneous data fields, and proposing precise repairs.

What Problem Does This Paper Address?

Auditing technical specifications is currently a fragmented process. Engineers often rely on a mix of expert-defined rules, statistical anomaly detectors, and manual cross-referencing with images. Purely "learned" models (like standard neural networks or LLMs) often struggle with the rigid physical laws—like lift-to-drag ratios or mass balances—that govern engineering. Conversely, traditional rule-based systems are too brittle to handle the "noise" and natural variations found in real-world fleet data.

The gap addressed here is the lack of an integrated, interpretable pipeline that can:

  1. Detect an inconsistency.
  2. Diagnose which physical law is being broken.
  3. Identify exactly which number in the spreadsheet is wrong (abduction).
  4. Suggest a corrected value (repair).
  5. Verify the claim against images (multimodal consistency).

Methods

The core of NeSySpec is Logic Tensor Networks (LTN), a neuro-symbolic framework that maps logical symbols onto tensors and differentiable functions.

Key innovations include:

  • Graded Predicates: Instead of "True" or "False," physical relations are treated as differentiable "graded" predicates. The system learns the "tolerances" (the acceptable bands of variation) for these laws from actual fleet data rather than relying on hard-coded expert bounds.
  • Abductive Search: To find the "altered field," the system performs an abductive search—essentially asking, "Which single change to this record would best restore logical consistency?"
  • Masked Stratified Aggregation: This allows the model to "abstain" from a verdict if the necessary data or image evidence is missing, preventing the "hallucinations" common in purely generative models.
  • Visual Grounding Axioms: A novel grounding axiom encourages the vision module to focus its "attention" on the specific parts of an aircraft (like engines or wings) that correspond to the data field being verified.
Figure 1: Overview of the NeSySpec pipeline.

Datasets

The authors validated NeSySpec using a robust multimodal setup:

  • Aircraft Specifications: A dataset of 107 aircraft types across 71 families, covering 16 numeric fields (mass, thrust, wingspan, etc.).
  • FGVC-Aircraft: 6,800 ground photographs used for visual verification.
  • RarePlanes: 14,707 annotated aircraft in satellite imagery.
  • Synthetic Benchmarks: To test the auditor, the authors generated "corruptions" (impossible values), "Frankenstein" swaps (mixing parts from different aircraft classes), and deliberate falsifications.

Key Insights and Ideas

  • Logic as a Diagnostic Tool: By encoding physics as logic, the model provides built-in explainability. When it flags a record, it can explicitly state, "This violates the Breguet range equation," making it a "glass-box" system.
  • Learning the "Normal": One of the "aha" moments is that the model learns physical tolerances from the data itself. It discovers what a "reasonable" thrust-to-weight ratio looks like for a narrow-body turbofan by looking at the fleet, rather than just following a textbook.
  • Abduction vs. Localization: The paper shows that abductive reasoning (searching for the cause of a logical failure) is significantly more accurate at finding errors than standard statistical localization methods like Fellegi-Holt.
  • Multimodal Synergy: The system can detect if a record says an aircraft has four engines while the satellite photo only shows two, naming the specific "cross-modal" axiom that was contradicted.

Key Results

NeSySpec was compared against a wide array of state-of-the-art (SOTA) methods, as summarized in the table below.

Table 2: Comparison with the state of the art.
  • Constraint Diagnosis: NeSySpec correctly named the violated constraint in 94.2% of corrupted test records, far outperforming the classical Fellegi-Holt method (79.5%).
  • Field Localization: It identified the altered field in 59.3% of records, nearly doubling the performance of minimum-cardinality localization (30.6%).
  • Multimodal Detection: On satellite imagery, it detected incoherent record-image pairs with a near-perfect AUROC of 0.994.
  • LLM Comparison: NeSySpec significantly outperformed prompted language models (like Qwen3-VL) in technical reasoning tasks, proving that specialized neuro-symbolic logic is still superior to general-purpose LLMs for high-precision engineering audits.

Limitations and Future Work

While powerful, NeSySpec has clear boundaries:

  • Domain Specificity: The current knowledge base is aircraft-specific. Porting it to other fields (like civil engineering or industrial equipment) requires defining new physical axioms and training new visual predicates.
  • Synthetic vs. Real Errors: The evaluation relied on synthetic errors. Testing against "naturally occurring" errors in messy, real-world databases is a necessary next step.
  • Generative Integration: The authors suggest that future versions could connect to agentic pipelines to generate parametric CAD models directly from audited specifications.

NeSySpec represents a significant step forward for "Technical Intelligence," showing that when it comes to auditing the physical world, a little bit of logic goes a long way.

Summary updated

Abstract

Engineers who audit a declared specification of an engineered system need to know which physical constraints it respects, which it violates, which field was altered and how to correct it. NeSySpec is an integrated and interpretable neuro-symbolic audit framework, built on Logic Tensor Networks, that answers these questions in one pipeline. Its axioms encode the physical relations of the domain as differentiable graded predicates whose tolerances are learned from a fleet of real records; from the same degrees, it diagnoses the violated constraint, identifies the altered field by abduction, proposes a minimum-change correction and ties image evidence to the parts it describes, abstaining exactly where evidence is missing, and returns all of it in one verification report. We validate NeSySpec on 107 aircraft types, with 16 test aircraft of 9 families held out of training, on specification records paired with ground photographs and satellite imagery and on synthetic errors: generated corruptions, edits labeled from the edit alone and deliberate falsifications. NeSySpec combines all the assessed audit outputs in one pipeline, which no learned, rule-based, solver-based, neuro-symbolic or language-model comparator does; on the test aircraft none is ahead of it in the macro-F1 of its constraint verdicts, in naming the violated constraint and the altered field or in the repair error, and it is competitive in record and image-pair detection. Its diagnosis names the violated constraint in 94.2% of corrupted test records, against 79.5% for classical Fellegi-Holt diagnosis, and on edits labeled from the edit alone, without retraining, abduction finds the altered field in 59.3% of records, against 30.6% for minimum-cardinality localization with random tie- breaking. Only the knowledge base, the visual predicates, the corruption generator and the data are specific to the domain.
BibTeX
@misc{basciano2026nesyspec,
  author = {Pier Mattia Basciano and Gianvito Antenori and Cristiano Andolfi and Giovanni Farina and Salvatore Monteleone},
  title = {{NeSySpec: Neuro-Symbolic Technical Intelligence for Auditing Engineering Specifications with an Application to Aircraft}},
  year = {2026},
  eprint = {2610.35403},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.35403},
}
Figure