LAVA is a research prototype for site-specific underground rock-bolt corrosion screening. It is intended for mining and tunnel operators, geotechnical engineers and engineering consultants exploring evidence-led review workflows.
These are the same reviewed answers used in the interactive preview. Each answer links to approved public sources.
Is the heat map a prediction of site corrosion risk?
No. The public dashboard interpolates eight synthetic endpoint severities over a fictional base map using inverse-distance weighting (IDW). The colours help a reviewer explore regional differences and select a case; they are not new model predictions, measured conditions between supports or probabilities of failure. A production risk map would require verified asset locations, inspection records, exposure and consequence criteria, and engineering review.
How does LAVA connect the underground environment to rock-bolt corrosion?
LAVA combines environmental and support characteristics to estimate separate corrosion-related material responses. The accepted method uses a site-specific chemistry score and a pre-assigned Rock Reactivity Index, alongside age, material and matched-control context. The public overview identifies chloride, copper and electrical conductivity as chemistry inputs, with pH retained as environmental context. These relationships are learned within the study scope; they are not a universal corrosion law or proof that the same calibration applies to another mine or tunnel.
For this example, keep the ductility flag open while assembling the evidence needed for review. Check dated water-quality records, support identity and inspections, the engineer-reviewed geological context and RRI assignment, and the site’s approved criteria. Then record the reviewer’s assessment and any justified follow-up. SYN-03 supplies none of those site histories, so it cannot prescribe an inspection interval or replacement decision. The stored 83.83% static and 58.70% ductility reserves remain unchanged by the explanation.
How does LAVA help with corrosion and maintenance?
LAVA combines machine learning with an AI agent. The models estimate corrosion-related changes in rock-bolt performance from environmental and material inputs. Copilot explains those results, brings together relevant records and documents, and identifies missing information. Engineers can use that evidence when considering where inspection or maintenance needs attention. More targeted work and better-informed risk review are the intended benefits; savings and risk reduction still need to be evaluated in a site-specific pilot.
SYN-03 is a synthetic CTGAN_0001 record at 437 days. Its stored static reserve is 83.83%, while ductility reserve is 58.70% (rounded from the source values). The static screen passes the 66% showcase threshold; the provisional 70% ductility sensitivity screen triggers review. Keep both branches visible: this is a review cue for a synthetic example, not a safety finding or failure probability.
The showcase has no dated water history, inspected support inventory, engineer-reviewed geological assessment, or approved company policy. Request records tied to the real support or site and retain their dates and provenance. The displayed region and shared-batch assumption do not identify an individual support or its installation date.
No. SYN-03 carries RRI 0.5 as a supplied input; it is not a newly verified geological judgement. A real review would need the relevant engineer-reviewed geological assessment and RRI assignment record. This synthetic row cannot establish a historical change or site condition.
They represent different response pathways. The static branch combines section integrity and ultimate-strength response; the ductility branch keeps elongation reserve visible. A relatively high static reserve can coexist with a more restricted ductility reserve. Showing each value and flag preserves which pathway is driving review.
The paper overview names XGBoost for the section-loss proxy and control-normalised total-elongation loss, and Gaussian-process regression for coupon ultimate-strength loss relative to matched controls. The serviceability index takes the more restrictive of the static-loss sum and elongation loss. These component learners have distinct jobs; the public overview does not disclose training recipes or fitted artefacts.
Are the displayed thresholds failure probabilities?
No. The 66% static screen and provisional 70% ductility sensitivity threshold are demonstration screening rules, not calibrated failure probabilities or universal material limits. A triggered branch routes an example to human review; a passing branch does not mean the support is safe.
No. It is a conditional input-response range under a selected showcase assumption. It is not a calibrated 95% prediction or confidence interval and does not include model error, parameter uncertainty, unvaried factors, or variation between real supports. Smoothing between stored query points adds no new model evaluations.
A bounded pilot could test whether approved records can be organised into a traceable review workflow: show stored or validated results, preserve separate branch flags, identify missing evidence, and attach readable sources. The current public demo can illustrate that interface flow. It does not establish field performance, production readiness, or operational approval.
A real evaluation would need authorised, appropriately governed site records, such as dated water-quality measurements, support identity and inspection history, relevant material/test records, and engineer-reviewed geological context. The public showcase contains none of these real operational records; missing inputs must remain explicit.
Assess whether reviewers can trace each displayed value to its evidence class and source, distinguish stored results from observations and assistant prose, see disagreements between branches, and identify gaps without losing flags. Any claim about predictive performance or operational value would require a separately designed and governed evaluation with suitable site evidence.
Can Copilot approve an asset or recommend replacement?
No. Copilot can explain approved material, identify evidence gaps, and draft a review summary. It cannot approve safety, modify stored results, issue an inspection or replacement order, or substitute for the responsible engineer and authorised company process.