PaperlayerView exact revision on arXiv

Source-linked AI summary

Generative artificial intelligence for reliable mechanistic reasoning for corrosion

Bharath M N, R K Singh Raman, Alankar Alankar

arXiv:2609.00099v1cs.LGcond-mat.mtrl-sci

TL;DR

Corrosion models can predict rates but often lack mechanistic reasoning, while existing factuality metrics miss errors in causal chains. This paper combines domain-adapted retrieval-augmented generation with Reason Map auditing, improving Token F1 by 143–194% and detecting unsupported mechanistic inferences.

  • Problem

    Existing corrosion models and factuality metrics provide limited mechanistic interpretability and cannot reliably assess the reasoning chains connecting generated claims.

    Source: §1 · p. 2

  • Method

    The study fine-tunes language models on expert-verified corrosion literature, adds hybrid retrieval, and audits generated reasoning with proposition-based evidence graphs.

    Source: §1 · p. 6, §4.6 · p. 23

  • Results

    Token F1 increased by 143–194% with retrieval augmentation, while Reason Map detected causal reversals and unsupported inferential leaps missed by flat factuality metrics.

    Source: §3 · p. 19, Conclusion · p. 24

  • Takeaways & Limitations

    Reliable corrosion knowledge synthesis requires evaluating both claim correctness and the mechanistic reasoning structure that connects claims.

    Source: Conclusion · p. 24

    The study focuses on magnesium alloy corrosion, with broader validation needed across alloy systems, degradation mechanisms, and exposure environments.

    Source: §3 · p. 20

Abstract

from arXiv · show

Corrosion accounts for approximately 4% of global GDP, and reliable prediction is essential for timely mitigation. Machine learning effectively predicts corrosion rates from composition, microstructure, and environmental variables, but cannot explain the underlying mechanisms. A reliable approach in safety-critical materials engineering requires not only accurate retrieval but also mechanistically defensible reasoning, a capability that existing factuality metrics cannot assess. This work presents a domain-adapted retrieval-augmented generation framework for corrosion knowledge synthesis, demonstrated on magnesium alloy corrosion. Three open-weight language models (Llama-3.1-8B, Qwen-2.5-7B, Mistral-7B) are fine-tuned on 3,309 expert-verified question-answer pairs from 840 peer-reviewed papers and integrated with a hybrid dense-lexical retrieval pipeline. Retrieval augmentation produces Token F1 gains of 143-194%, with system faithfulness of 0.964 and context recall of 0.988. Blind external validation on newly published literature and in-house electrochemical data confirms trend-level generalisation. Reason Map, a proposition-graph framework, is further introduced; it independently constructs directed evidence graphs from generated answers and retrieved literature, enabling systematic detection of causal direction inversions and unsupported inferential leaps that flat factuality metrics cannot expose. The modular architecture can be applied across domains, offering a generalizable blueprint for trustworthy AI-assisted knowledge synthesis to circumvent corrosion, which can also be applied to other engineering domains.

1. Corrosion assessment in the age of data-driven methods: from pattern recognition to mechanistic reasoning · 2. Results

This section presents and evaluates a retrieval-augmented framework for corrosion knowledge synthesis, covering generation quality, grounding, external generalisation, and mechanistic reasoning. Reason Map extends validation beyond claim-level factuality by identifying supported relations, unsupported inferential leaps, and causal-direction reversals.

1. Corrosion assessment in the age of data-driven methods: from pattern recognition to mechanistic reasoning

  • Corrosion assessment requires mechanistic synthesis beyond pattern-based prediction because interacting composition, microstructure, chemistry, and exposure variables limit conventional methods and narrow-data ML.RAG enables evidence-grounded synthesis across dispersed corrosion literature, but retrieved evidence alone does not guarantee correct causal reasoning.Source: §1 · pp. 1–3

2.1 Domain-adapted language models as generation baselines

  • Without retrieval, Llama-3.1-8B was the strongest baseline across all metrics, achieving Token F1 0.635, ROUGE-L 0.581, BLEU 0.413, and BERTScore F1 0.924.Mistral and Qwen retained moderate semantic alignment but had substantially lower lexical scores, indicating difficulty reproducing precise technical content.Source: §2.1 · p. 6, Figure 2 · p. 7

2.2 Retrieval augmentation transforms factual grounding across all three architectures

  • RAG improved all three models, with Token F1 gains of 143–194%; Mistral achieved the highest overall RAG performance, while Llama led BLEU.Mistral reached Token F1 0.737, ROUGE-L 0.740, and BERTScore F1 0.951; Llama led BLEU at 0.480.Source: §2.2 · p. 8, Figure 4 · p. 9

2.3 Automated and expert evaluation reveals reliable grounding and model-specific trade-offs

  • Automated and blinded expert evaluations showed reliable grounding, with Llama leading expert overall scores at 3.93/5 and Mistral leading answer correctness at 0.920.Content completeness and practical relevance remained weaker than technical accuracy across models.Source: §2.3 · pp. 10–11, Figure 7 · p. 12

2.4 Out-of-distribution validation confirms trend-level generalisation with residual weaknesses in fine-grained mechanistic fidelity

  • On unseen literature and electrochemical data, Llama performed best overall, scoring 24/36 and 7/9, while all models handled broad trends better than fine-grained mechanisms.Models struggled particularly with dataset-specific inflection points and distinguishing kinetic stabilisation from true passivation.Source: §2.4 · pp. 13–14, Figure 8 · p. 13, Figure 9 · p. 14

2.5 Reason Map detects reasoning errors missed by claim-level evaluation

  • Reason Map exposed causal reversals and unsupported inferential leaps despite high claim-level support, including a rare-earth case with faithfulness 0.964, seven supported nodes, and one reversed edge.A second hydrogen-assisted cracking case contained two causal reversals and two unsupported leaps; self-correction removed factual errors but not reversed dependencies.Source: §2.5 · pp. 15–17, Figure 10 · p. 16
  • Answer-graph validation distinguishes valid evidence paths, unsupported leaps, and causal reversals that claim-level evaluation misses.Valid edges include P1→P2→P4→P5, while P3→P2 and P3→P5 are unsupported and P5→P1 and P4→P1 reverse the supported direction.Source: Figure 11 · p. 18
  • 0.994, 0.996, and 0.991 entailment scores strongly anchor propositions about SCC mechanisms, WE43 hydrogen embrittlement, and Sr-HA coating suppression of H₂ evolution.AZ91D propositions on hydrogen accumulation and embrittlement receive NEI entailment scores of 0.073 and 0.074.Source: Figure 11 · p. 18
  • Reason Map uses an independently derived evidence graph as the reference structure for validating directed edges in generated answer graphs.The evidence graph contains hydrogen-embrittlement and hydrogen-evolution subnetworks, including E1–E4, E7–E8 and E5→E6→E9→E10.Source: Figure 12 · p. 18

3. Discussion

Retrieval augmentation substantially improves corrosion-model reliability but does not ensure mechanistically sound reasoning, motivating Reason Map’s structural evaluation of causal chains. The study also identifies validation limitations and a path toward training-time reasoning alignment and broader multimodal, cross-domain deployment.

Reliability and model trade-offs

  • Retrieval augmentation increased Token F1 by 143–194% and reduced grounding gaps from approximately 0.60 to 0.21–0.35, but did not eliminate them.Mistral achieved the strongest factual accuracy, Llama the most balanced performance, and Qwen faster responses but weaker mechanistic consistency, so deployment priorities determine model choice.Source: §3 · p. 19

Reasoning-structure evaluation

  • Reason Map detected causal reversals and unsupported inferential links that persisted despite high RAGAS faithfulness and claim verification.These structural errors can change corrosion recommendations, such as whether to target film nucleation mode or density and how to interpret stable open-circuit potential.Source: §3 · p. 19

Reasoning alignment

  • Post-generation correction reduced factual errors, but reversed causal dependencies remained, indicating that reasoning structure is harder to repair than unsupported content.The results motivate using Reason Map signals, unsupported leaps, causal reversals, and NEI rates as training-time alignment objectives rather than only post-hoc audits.Source: §3 · pp. 19–20

Limitations

  • The study’s scope is limited by magnesium-alloy focus, six external-validation examples, a compact dataset with four timepoints, one alloy, and one electrolyte, and a 140-question test set.The held-out test set showed Cohen's d = 2.1–3.1 but was underpowered for definitive conventional significance testing; broader multimodal and environmental validation is needed.Source: §3 · p. 20

Future directions and transferability

  • The framework provides a transferable foundation for transparent, evidence-grounded materials-science AI through hybrid retrieval, open-weight adaptation, external and experimental validation, and proposition-graph reasoning checks.Future directions include temporally updated benchmarks, multimodal corrosion intelligence, and applying the approach to alloy design, degradation modelling, heterogeneous catalysis, and biomaterials engineering.Source: §3 · p. 20

4. Methods

The methods combine an expert-curated magnesium-corrosion corpus with parameter-efficient model fine-tuning, hybrid retrieval, multi-level evaluation, and proposition-graph auditing. External literature and withheld electrochemical data assess generalisation, while Reason Map evaluates whether generated causal reasoning is evidence-supported.

4.1 Dataset construction and curation

  • The study assembled 840 open-access magnesium-corrosion papers, extracted 99.9% successfully with GROBID 27, and produced 3,309 expert-verified Q-A pairs split by document into train, validation, and test sets.The final partitions contained 2,604 training, 350 validation, and 355 test pairs; failed extraction was handled with PyMuPDF 28, with no documents lost.Source: §4.1 · pp. 20–21

4.2 Model selection and parameter-efficient fine-tuning

  • Three open-weight 7–8B instruction-tuned models were fine-tuned identically using LoRA 32 and QLoRA 34, updating approximately 0.5% of parameters with 4-bit base weights.The models were Llama-3.1-8B-Instruct, Qwen-2.5-7B-Instruct, and Mistral-7B-Instruct-v0.3; LoRA used rank 16, scaling factor 32, and dropout 0.1.Source: §4.2 · p. 21

4.3 Knowledge base construction and hybrid retrieval pipeline

  • The knowledge base combines SPECTER2 and BGE-M3 dense embeddings with FAISS and BM25 lexical indices, fusing 240 candidates by RRF before reranking and diversity-constrained selection.Paper and Q-A chunks preserve document metadata for source attribution; reranking uses BGE-M3, with at most two passages per document, paper-source inclusion, and MMR λ = 0.6.Source: §4.3 · p. 22

4.4 RAG evaluation protocol

  • RAG was compared with a no-RAG baseline on 140 held-out questions using five RAGAS dimensions and blinded expert ratings of technical accuracy, completeness, and relevance.Both conditions used greedy decoding with a maximum of 128 generated tokens; significance used 10,000 paired bootstrap resamples with two-sided α = 0.05, and practical significance used Cohen's d.Source: §4.4 · pp. 22–23

4.5 Out-of-distribution and experimental validation

  • Generalisation was tested on six post-cutoff literature questions and three undisclosed Mg–Ca/HBSS electrochemical queries scored for trends, mechanisms, directional accuracy, and quantitative consistency.The electrochemical measurements covered OCP over 3,600 s and impedance from 100 kHz to 0.01 Hz at 30 minutes, 1 day, 5 days, and 7 days.Source: §4.5 · p. 23

4.6 Reason Map: proposition-graph reasoning validation

  • Reason Map independently aligns answer and evidence proposition graphs to detect unsupported inferential leaps and causal reversals, with threshold-triggered retrieval, regeneration, and targeted correction.It classifies edges as valid, unsupported-leap, or causal-reversal and exports node- and edge-level verdicts, retry counts, and revision status.Source: §4.6 · pp. 23–24

Conclusion

The study shows that retrieval-augmented language models improve the factual grounding and reliability of corrosion knowledge synthesis when evaluation also examines mechanistic reasoning. Reason Map detects causal and structural reasoning failures that flat factuality metrics miss, while the modular approach is designed for extension beyond corrosion.

  • Retrieval augmentation improved the reliability and factual grounding of corrosion-specific answers across automated metrics, expert assessment, blind literature tests, and independent electrochemical experiments.The magnesium alloy corrosion application demonstrates that grounding generated responses in corrosion literature, technical evidence, and experimental context improves unguided parametric generation.Source: Conclusion · p. 24
  • Reason Map audited how corrosion evidence was organized into mechanistic chains, exposing causal reversals, unsupported inferential leaps, and misordered explanations missed by flat factuality metrics.These failures involve electrochemical reaction direction and sequence, oxide-film formation, hydrogen evolution, transport processes, and degradation pathways relevant to mitigation and design decisions.Source: Conclusion · p. 24
  • The combination of hybrid retrieval, multi-model fine-tuning, multi-paradigm evaluation, and proposition-graph reasoning is domain-agnostic and extendable to other materials-science applications.Proposed extensions include alloy design, degradation modelling, heterogeneous catalysis, and biomaterials engineering, supporting evaluation of reasoning structure alongside claims where errors have real consequences.Source: Conclusion · pp. 24–25

Supplementary document

The supplement documents model architectures, configurations, evaluation materials, and supporting corpus analyses. Blind validation shows partial success on mechanistically non-trivial magnesium-corrosion questions, with performance varying across SCC, corrosion-current, and time-scale reasoning.

Supplementary document

  • A supplementary corrosion-model example reports average R² = 0.78 ± 0.06 for a process-aware model versus 0.61 ± 0.04 for a prior simple DNN, with MAE of approximately 150 mV.Outliers were linked to missing processing-history text, while optimisation could overshoot plausible compositions.Source: Supplementary document · p. 28
  • Blind evaluation used six mechanistically non-trivial questions from three recently published magnesium-alloy corrosion papers excluded from training and indexing.The supplementary validation table scores environmental classification, trend ranking, and mechanistic explanation, with a maximum of 36 points overall.Source: Table S6 · p. 30, Supplementary document · p. 38

Example 1- SCC vs PDP contradiction (external paper validation)

  • Example 1 achieved scores of 5/6, 4/6, and 5/6, with models generally capturing higher icorr in BSA-containing conditions but inconsistently classifying SCC versus mechanical-overload outcomes.The strongest responses linked reduced SCC with protective fresh crack-tip films and increased long-exposure corrosion with chelation or complex formation that disrupts corrosion films.Source: Example 1- SCC vs PDP contradiction… · pp. 38–39

Example 2- Time-scale reasoning (minutes vs days)

  • Example 2 scores of 4/6, 3/6, and 1/6 show that models often identified early protection followed by late acceleration but frequently missed fresh-film SCC control and chelation-driven film disruption.The lowest-scoring response incorrectly predicted worsening SCC and omitted the paper’s fresh-crack-tip-film and complex-formation mechanisms.Source: Example 2- Time-scale reasoning… · pp. 39–40

Example-3: AZ63 vs AZ91 corrosion paradox

Across five as-cast AZ alloys in 3.5 wt.% NaCl at 25 °C, AZ31 is most corrosion-resistant and AZ63 least, while AZ91 is most localized and AZ63 most uniformly aggressive. The paradox is explained by eutectic α buffering galvanic attack in AZ63 but greater corrosion-product and β-phase desorption, whereas AZ91 retains products and forms an anodic barrier network.

Gold answer

  • AZ31 > AZ61 > AZ62 > AZ91 > AZ63 ranks overall corrosion resistance, with AZ91 most localized and AZ63 most uniform yet fastest-corroding.Immersion and electrochemical results agree: AZ31 has the lowest corrosion rate, AZ63 the highest, and AZ91 is worse than AZ62 but not the fastest overall.Source: Example-3: AZ63 vs AZ91 corrosion… · p. 41
  • AZ63’s greater eutectic α buffers severe β/α micro-galvanic attack and promotes uniformity, but desorption of corrosion products and discrete β phases exposes fresh Mg surfaces for continued attack.In AZ91, the larger β-Mg17Al12 network instead provides an anodic barrier effect and tends to retain corrosion products, producing severe localization without the highest overall rate.Source: Example-3: AZ63 vs AZ91 corrosion… · p. 41

Model responses

  • Model answers disagreed with the gold ranking: Llama and Qwen identified AZ63 as fastest, but Qwen ranked AZ91 worst and Mistral produced a different AZ31 > AZ62 > AZ61 > AZ63 > AZ91 order.The responses generally identified AZ91 as most localized and AZ63 as uniformly fastest, but several explanations incorrectly assigned phase effects or corrosion behavior.Source: Example-3: AZ63 vs AZ91 corrosion… · p. 41

LLAMA — Score: 3/6

Llama scored 3/6 because it partially identified corrosion-rate trends and mechanisms but confused alloy chemistry and failed to resolve the AZ63–AZ91 localization paradox.

LLAMA — Score: 3/6

  • Llama scored 3/6, correctly identifying AZ31 as best and AZ63 as very poor or fast, but incorrectly making AZ91 worst and attributing AZ91 to high Zn.It also reversed the β-Mg17Al12 comparison, stating that AZ63 contains more β phase than AZ91.Source: LLAMA — Score: 3/6 · p. 42
  • The answer only partially distinguished AZ63 as fastest, AZ91 as most localized, and AZ63 as most uniformly fast.It mentioned eutectic α preferential corrosion and fresh-surface exposure but did not correctly explain the AZ63–AZ91 paradox.Source: LLAMA — Score: 3/6 · p. 42

QWEN — Score: 4/6

Qwen scored 4/6 by capturing key AZ91/AZ63 corrosion trends while misranking AZ63 and AZ61/AZ62 and misstating the Zn-driven mechanism. In AZ91 EIS reasoning, the gold answer distinguishes three loops at free corrosion and anodic polarization from one cathodic loop through anodic-state variables.

QWEN — Score: 4/6

  • Qwen scored 4/6, correctly identifying AZ91 as most localized and AZ63 as uniform but fastest, but misranking AZ63 and AZ61/AZ62.It also missed that eutectic α, rather than eutectic β, is the crucial Zn-driven feature behind AZ63 behavior.Source: QWEN — Score: 4/6 · p. 42

Example 4: AZ91 EIS state-variable reasoning under OCP vs ±120 mV polarization

  • At free corrosion and +120 mV, AZ91 shows two capacitive loops plus one inductive loop, whereas −120 mV leaves only one capacitive loop.The EIS response combines anodic and cathodic processes near free corrosion, but strong cathodic polarization suppresses anodic dissolution and removes the extra responses.Source: QWEN — Score: 4/6 · pp. 42–43
  • The gold mechanism assigns the extra responses to anodic variables θ and CMg+, with θ capacitive and CMg+ inductive contributions disappearing under strong cathodic polarization.Cathodic polarization suppresses the anodic pathway to Mg+ and makes θ potential-invariant, leaving the double-layer capacitive loop.Source: QWEN — Score: 4/6 · p. 43
  • Qwen reproduced the loop-count trend but incorrectly described free-corrosion AZ91 as having three capacitive loops and misassigned the surviving cathodic response.Mistral correctly retained the inductive loop under anodic polarization, while Llama incorrectly attributed the cathodic change to θ = 1 and suppressed CMg+.Source: QWEN — Score: 4/6 · p. 43

Mistral — Score: 1/6

Mistral scored 1/6, correctly handling the free-corrosion and strong-cathodic cases but failing on anodic loop persistence, variable assignments, and the cathodic mechanism.

Mistral

  • 1/2: Mistral correctly identifies the free-corrosion case and one remaining capacitive loop under strong cathodic polarization, but wrongly removes the inductive loop under anodic polarization.The inductive loop persists at +120 mV.Source: Mistral — Score: 1/6 · p. 43
  • 0/2: Mistral mentions θ and CMg+ but does not correctly assign θ to the capacitive response and CMg+ to the inductive response.Source: Mistral — Score: 1/6 · p. 43
  • 0/2: Mistral incorrectly attributes disappearance of the extra loops to hydroxide consumption or pH reduction rather than loss of anodic-state-variable responses under strong cathodic polarization.Source: Mistral — Score: 1/6 · p. 44

Llama — Score: 3/6

Llama correctly identifies the loop-pattern changes across corrosion conditions but misattributes the underlying electrochemical elements and overstates the cathodic-polarization interpretation.

Loop-pattern recognition

  • Llama scores 2/2 for correctly identifying two capacitive plus one inductive loop at free corrosion and +120 mV, versus one capacitive loop at −120 mV.The loop pattern changes with polarization condition, while the specific loop attribution is evaluated separately.Source: Llama — Score: 3/6 · p. 44

Loop attribution

  • Llama scores 0/2 because it incorrectly attributes AZ91 responses to inner film, outer film, and double layer, missing θ → capacitive and CMg+ → inductive mapping.The intended state-variable mapping is θ to capacitive behavior and CMg+ to inductive behavior.Source: Llama — Score: 3/6 · p. 44

Cathodic-polarization interpretation

  • Llama scores 1/2 for recognizing CMg+ suppression under strong cathodic polarization but incorrectly claiming complete corrosion-product coverage and disappearance of the double-layer response.The remaining loop is identified as the double-layer capacitive response, not its disappearance.Source: Llama — Score: 3/6 · p. 44

Qwen — Score: 1/6

Qwen scored 1/6 because it misinterpreted corrosion-model loop assignments, state variables, and cathodic suppression. In a separate blind test, Qwen scored 2/6 by identifying the best condition but misranking conditions and proposing an unsupported mechanism.

Qwen — Score: 1/6

  • Qwen scored 1/6, misreading free-corrosion and anodic spectra as three capacitive loops instead of two capacitive and one inductive loop.It correctly recognized that strong cathodic polarization leaves one capacitive loop.Source: Qwen — Score: 1/6 · p. 44
  • Qwen scored 0/2 on state variables and cathodic mechanism, missing that θ gives the capacitive response, CMg+ gives the inductive response, and polarization suppresses the anodic Mg+ pathway.Source: Qwen — Score: 1/6 · p. 44

Example 5: Finest grains do not equal best corrosion resistance

  • The corrosion optimum reflects precipitate homogenization, solute re-solution into α-Mg, reduced micro-galvanic heterogeneity, and stable alloy-oxide formation rather than grain size alone.For the film comparison, BM forms a thicker porous Mg(OH)2-rich layer, whereas FSP 1200–200 forms a thinner compact film with higher Rct and Rp.Source: Example 5: Finest grains do not equal… · pp. 45, 47
  • Qwen scored 2/6 in Example 5, correctly identifying 1200–200 as best but wrongly assigning the finest grains to it and misranking the intermediate conditions.The paper supports 1200–100 as finest and the corrosion-resistance order BM → 1200–50 → 1200–100 → 1200–200.Source: Example 5: Finest grains do not equal… · pp. 45–46

Llama — Score: 5/6

Llama scored 5/6, correctly distinguishing corrosion-product thickness and film protection while making one error concerning Rf after 24 h.

Qwen — Score: 4/6

Qwen scored 4/6, correctly identifying the central electrochemical pattern and final mechanism but reversing which sample forms the thicker corrosion-product layer.

  • Qwen reverses the core thickness result by claiming FSP, rather than BM, forms the thicker corrosion-product layer.The paper shows that BM forms the thicker layer.Source: Qwen — Score: 4/6 · p. 48
  • 4/6: Qwen correctly assigns higher Rf to BM and higher Rct/Rp to FSP, matching the central electrochemical pattern.These assignments were judged fully correct.Source: Qwen — Score: 4/6 · p. 48
  • Qwen’s final mechanism is strong: thicker products are not necessarily more protective, and porous Mg(OH)2-rich products underperform compact RE-oxide/alloy-oxide-enriched films.The mechanism aligns with the paper, although it is paired with the wrong sample assignment in the first bullet.Source: Qwen — Score: 4/6 · p. 48

Mistral — Score: 0/6

Mistral scored 0/6, incorrectly identifying corrosion-product thickness, electrochemical resistance assignments, and the governing protection mechanism.

  • Mistral scored 0/6 overall, failing all three evaluated criteria.It incorrectly assigned corrosion-product thickness and electrochemical resistance values, while inverting the proposed corrosion mechanism.Source: Mistral — Score: 0/6 · p. 48
  • A = 0/2: Mistral incorrectly assigned the thicker corrosion-product layer to FSP rather than BM.Source: Mistral — Score: 0/6 · p. 48
  • B = 0/2: Mistral wrongly assigned Rct and Rp to BM, although the paper reports both are much higher for FSP 1200–200.Source: Mistral — Score: 0/6 · p. 48
  • C = 0/2: Mistral inverted the mechanism, attributing the more protective alloy-oxide/RE-oxide-enriched film to BM rather than FSP.Source: Mistral — Score: 0/6 · p. 48

Supplementary Note S2: Experimental validation using electrochemical datasets

This section validates mechanistic predictions against electrochemical datasets for Mg-Ca in protein-free HBSS at 37°C, focusing on low-frequency impedance and OCP behavior. The observations characterize active, dynamic corrosion despite OCP stabilization and relate inductive responses to hydrogen-evolution intermediates, film formation and breakdown, and possible negative difference effects.

Low-frequency impedance

  • Chloride attack was consistently linked to Mg(OH)₂ dissolution and reduced impedance, while calcium-phosphate precipitation was linked to partial pore sealing or increased protection.Qwen additionally proposed MgCa₂(PO₄)₂ and CaHPO₄ deposition, followed by mixed-film stabilization through 5 days.Source: Supplementary Note S2: Experimental… · pp. 49–50

Open circuit potential

  • OCP predictions described an initial shift during oxide transformation, a more protective mixed Mg(OH)₂/calcium-phosphate film by 1 day, and anodic movement as chloride penetrates and degrades the film.Llama associated the stages with anodic, cathodic, and subsequent anodic shifts, respectively.Source: Supplementary Note S2: Experimental… · p. 51
  • At 7 days, the models associated stabilized, strongly negative OCP with active corrosion rather than passivation, reflecting balanced Mg dissolution, hydrogen evolution, and film formation or dissolution.Qwen gave a stabilized range of -1.55 to -1.65 V vs SCE.Source: Supplementary Note S2: Experimental… · p. 51
  • Qwen predicted that OCP rapidly rises toward nobler potentials during the first 30 minutes in HBSS, then becomes significantly nobler than the bare alloy (~-1.55 V vs SCE) as protective films form.Over days 1–5, OCP initially drifts nobler due to film growth, then fluctuates as Cl⁻ penetrates.Source: Supplementary Note S2: Experimental… · p. 51

Q3: EIS interpretation

  • Inductive behavior is attributed to adsorbed magnesium and hydrogen-related intermediates associated with the negative difference effect and hydrogen evolution.The model responses identify Mg²⁺, Mg(OH)⁺, Mg⁺, and adsorbed hydrogen species as contributors to the low-frequency inductive response.Source: Supplementary Note S2: Experimental… · p. 53
  • Persistent inductive loops through 7 days show that the corrosion-product layer does not achieve true passivity.A passive film would suppress or eliminate inductive behavior, whereas the observed loop indicates ongoing electrochemical activity and hydrogen evolution.Source: Supplementary Note S2: Experimental… · p. 53
  • A dense protective Ca-phosphate film would introduce a capacitive film branch and eliminate or substantially diminish the inductive loop.The predicted spectral change contrasts with the observed persistent inductive response and minimal capacitive arc.Source: Supplementary Note S2: Experimental… · p. 53
  • Fewer than 5 of 81 EIS points show capacitive character, indicating that inductive relaxation dominates double-layer and film capacitance.This compressed capacitive response reflects a dynamic, non-protective corrosion-product layer rather than a dense protective film.Source: Supplementary Note S2: Experimental… · pp. 53–54

Q4: Interpreting stable OCP

  • −1.834 V OCP, <1 mV drift over 3,600 seconds, and ~3,310 Ω impedance at 0.01 Hz collectively indicate stable active corrosion, not passivation.The 7-day OCP was the most negative recorded, the impedance was the lowest across timepoints, and the dominant inductive loop remained consistent with hydrogen evolution.Source: Supplementary Note S2: Experimental… · pp. 54–55
  • The stable OCP, low impedance, and persistent inductive response instead indicate a dynamic corrosion state in which products and hydrogen evolution continually destabilize the surface film.This combination directly contradicts the claim that OCP stabilization alone demonstrates passivation.Source: Supplementary Note S2: Experimental… · pp. 54–55

Electrochemical evidence

  • At 7 days, OCP was −1.834 V and low-frequency impedance was 3,310 Ω, while an inductive loop appeared at 0.01 Hz.These observations differ from the expected impedance increase during passivation.Source: Supplementary Note S2: Experimental… · p. 55

Mechanistic interpretation

  • The stable OCP and low impedance are consistent with a metastable state of rapid protective-film formation and breakdown, producing active dissolution and partial protection.Continuous charge transfer may stabilize OCP, while film instability maintains low impedance.Source: Supplementary Note S2: Experimental… · p. 55
  • The strongly negative OCP and 3,310 Ω impedance, compared with earlier 10⁴-10⁵ Ω values, indicate active corrosion rather than passivation.A stable OCP alone does not establish passivation when the potential remains strongly negative.Source: Supplementary Note S2: Experimental… · p. 55
  • Persistent inductive behavior is attributed to adsorption-desorption processes and NDE, in which hydrogen evolution creates cathodic sites that destabilize protective films.Together, stabilized active-potential OCP, low impedance, and the inductive loop indicate NDE rather than passivation.Source: Supplementary Note S2: Experimental… · p. 55

Related papers

Loading 2609.00099v1…