Digital Geotechnics, AI, Ethics and Future Practice

Introduction and context

Geotechnical engineering is becoming a data-intensive discipline. Conventional investigations remain essential, but modern projects increasingly combine in situ sensors, automated data acquisition, geographic information systems (GIS), building information modelling (BIM), remote sensing, machine learning and digital twins. These technologies can improve the observation of ground behaviour, support earlier warning of failure, reduce uncertainty and improve whole-life asset management. They do not, however, remove the need for engineering judgement, appropriate ground models or professional accountability.

Digital geotechnics may be defined as the systematic use of digital data, computational models and connected information systems to investigate, analyse, design, construct and monitor geotechnical assets. Its central challenge is not simply collecting more data. It is converting heterogeneous observations into reliable decisions while preserving traceability, explainability, security and ethical responsibility.

A digital system is only as credible as its weakest component. A highly sophisticated machine-learning model cannot compensate for poorly located instruments, unrepresentative samples, sensor drift, inappropriate boundary conditions or biased training data. Accordingly, digital transformation should be understood as an extension of the observational method rather than a replacement for soil mechanics.

Learning objectives

By the end of this chapter, the reader should be able to:

1. Explain the architecture and engineering value of digital geotechnical systems.

2. Select and critically evaluate sensors, monitoring networks and remote-sensing methods.

3. Distinguish the roles of GIS, BIM, digital twins and geotechnical information models.

4. Describe appropriate uses and limitations of machine learning and data-driven soil models.

5. Evaluate uncertainty, explainability, bias, cybersecurity and data-governance risks.

6. Apply ethical and professional principles to automated geotechnical decisions.

7. Propose a responsible implementation strategy for future geotechnical practice.

From conventional investigation to digital geotechnics

A conventional ground investigation produces borehole logs, laboratory results, in situ tests, geological interpretation and design parameters. Digital practice adds continuous observations, spatial databases, automated processing and links between design, construction and operation. The resulting workflow can be viewed as a chain:

observation → transmission → storage → quality control → interpretation → prediction → decision → feedback.

Failure at any stage can produce misleading conclusions. For example, a piezometer may measure pore pressure accurately, yet a blocked filter, incorrect datum or faulty timestamp can make the resulting trend unusable. Similarly, a model may predict settlement well within the calibration range but fail when applied to a soil fabric, stress path or drainage condition not represented in its training data.

Important terminology includes:

• Data: recorded observations, measurements or derived values.
• Information: data interpreted in a meaningful context.
• Model: a simplified representation used to describe, predict or explain a system.
• Digital thread: traceable connection among data, decisions, models and project stages.
• Interoperability: the ability of systems to exchange and correctly interpret information.
• Uncertainty: incomplete knowledge or variability affecting a quantity or decision.
• Observational method: a design approach in which behaviour is monitored and predefined actions are taken as evidence develops.

Digital systems are especially valuable where conditions change with time, such as deep excavations, embankment construction, tailings facilities, landslides, tunnels, offshore foundations and ageing infrastructure.

Sensor networks and automated monitoring

Sensors used in geotechnical engineering

A monitoring programme must be designed from the failure mechanisms and decisions it is intended to support. Typical instruments include:

• piezometers for pore-water pressure;
• inclinometers and in-place inclinometers for lateral deformation;
• extensometers and convergence meters for displacement;
• settlement plates, hydrostatic levelling systems and precise levelling points;
• load cells and pressure cells for structural or contact loads;
• vibrating-wire strain gauges and fibre-optic sensors;
• accelerometers for vibration and dynamic response;
• weather stations for rainfall, temperature and evaporation;
• GNSS receivers, robotic total stations and radar for surface movement.

The choice depends on measurand, range, resolution, accuracy, sampling frequency, environmental durability, installation disturbance, power supply, communications and maintenance requirements. Accuracy is not equivalent to resolution: a device may display small increments while having a substantial systematic error.

Network architecture

An automated monitoring system normally contains sensing nodes, a local data logger, communications, a server or cloud platform, analytics and a user interface. Critical systems should include independent power, time synchronisation, local buffering during communication loss, redundant communication paths and alarms that remain available during server outages.

Monitoring should be linked to action levels. A useful scheme has:

• a normal or green level, indicating expected behaviour;
• an alert or amber level, requiring verification and increased observation;
• an action or red level, requiring a predefined engineering response.

Thresholds should not be selected solely from arbitrary numerical limits. They should reflect design predictions, rates of change, uncertainty, consequences of failure and the time required to intervene. A rate-of-change alarm may be more informative than an absolute displacement alarm in a rapidly developing failure.

Data quality and sensor uncertainty

A credible monitoring plan includes calibration, installation records, baseline readings, validation checks, metadata, maintenance schedules and procedures for missing or anomalous data. Quality assurance should examine:

• range and plausibility checks;
• sudden jumps and flat-line signals;
• duplicated timestamps;
• sensor drift and hysteresis;
• cross-checks against independent instruments;
• environmental and temperature effects;
• changes caused by construction activities.

A simple measurement model is

\[

y = x + b + \varepsilon,

\]

where \(y\) is the recorded value, \(x\) is the physical quantity of interest, \(b\) is systematic bias and \(\varepsilon\) is random measurement error. The equation is a conceptual model, not a correction formula. It is useful because calibration can reduce uncertainty in \(b\), while repeated observations may reduce the influence of random error. Neither process eliminates errors caused by incorrect installation or an inappropriate interpretation of what the sensor measures.

Sensor placement should be based on expected mechanisms and spatial variability. Dense instrumentation is not automatically superior: poorly located sensors can create false confidence, whereas a smaller network designed around critical mechanisms may be more informative.

GIS, BIM and geotechnical information management

GIS

GIS stores, analyses and visualises spatially referenced information. In geotechnics, it can combine geology, boreholes, groundwater, topography, land use, hazards, infrastructure and remotely sensed deformation. GIS supports spatial interpolation, terrain analysis, route selection, landslide susceptibility mapping and regional asset screening.

Interpolation requires caution. A contour or raster surface is a model, not a direct observation. Results depend on sampling density, anisotropy, trend, support volume and the assumed spatial correlation structure. Extrapolation beyond the investigation domain is particularly hazardous. Geological boundaries should not be treated as smooth statistical surfaces when field evidence indicates discontinuities or abrupt changes.

BIM and geotechnical information models

BIM provides a structured information environment in which geometry, attributes, specifications, construction sequencing and asset information can be coordinated. A geotechnical model may contain strata, interpreted surfaces, ground investigation points, retaining elements, foundations, groundwater conditions and design assumptions.

The principal benefits are improved coordination, clash detection, change management and transfer of information into construction and operation. Common difficulties include incompatible data schemas, inconsistent coordinate systems, different levels of detail, uncertain ownership of interpreted geology and loss of metadata during file conversion.

A borehole log should not be represented merely as a coloured line. Its source, elevation datum, sampling method, laboratory basis, interpretation status and confidence should remain attached to the data. Separating observed facts from interpreted surfaces is fundamental to responsible information management.

Remote sensing and Earth observation

Remote sensing provides measurements without direct physical contact. Relevant methods include satellite and aerial optical imagery, LiDAR, photogrammetry, interferometric synthetic aperture radar (InSAR), terrestrial laser scanning and unmanned aerial systems.

InSAR can detect surface displacement over broad areas by comparing radar phase information acquired at different times. It is valuable for regional subsidence, landslides, embankments and infrastructure corridors. Interpretation is affected by vegetation, atmospheric disturbances, decorrelation, line-of-sight geometry and reference-point selection. It measures displacement in a radar line-of-sight direction, not automatically the full three-dimensional displacement vector.

LiDAR and photogrammetry produce dense surface models useful for erosion, deformation, excavation quantities and geomorphological mapping. Their accuracy depends on control points, surface texture, vegetation, viewing geometry and processing. Remote sensing should therefore be integrated with ground control and geotechnical instrumentation rather than treated as a complete substitute for it.

Digital twins and the observational method

A digital twin is a continuously updated digital representation of a physical asset or system, connected to observations and capable of supporting analysis or decisions. A three-dimensional BIM model alone is not necessarily a digital twin. A twin requires an explicit relationship between the physical system, data streams, models, update processes and intended decisions.

A geotechnical digital twin may include:

1. a spatial representation of ground and structures;

2. live or periodic monitoring data;

3. numerical or reduced-order models;

4. construction and maintenance records;

5. uncertainty and confidence information;

6. decision rules and scenario tools.

Digital twins are useful for comparing predicted and observed behaviour, updating risk assessments, testing intervention scenarios and preserving an asset’s information through its life cycle. Their limitations include model discrepancy, incomplete observability, high maintenance cost, cyber risk and the danger that users mistake visual sophistication for predictive validity.

The observational method provides a sound conceptual basis. Before construction, the engineer defines expected behaviour, measurable parameters, trigger levels and contingency actions. During construction, observations are compared with predictions and the design is adjusted within predefined limits. Digital systems improve the speed and scale of this feedback, but they do not remove the need to define credible mechanisms and actions in advance.

Machine learning and data-driven soil models

What machine learning contributes

Machine learning (ML) estimates relationships from data. Supervised learning uses labelled inputs and outputs, such as soil index properties and measured compressibility. Unsupervised learning identifies patterns or clusters, such as groups of similar ground conditions. Time-series methods detect trends or forecast sensor values. Classification models may identify likely geological units or failure states.

Potential applications include:

• estimating engineering parameters from investigation data;
• classifying soil or rock units;
• detecting anomalous sensor behaviour;
• forecasting settlement or pore pressure;
• prioritising inspection and maintenance;
• mapping landslide susceptibility;
• supporting inverse analysis and model calibration.

ML is most defensible when it complements mechanics. A hybrid model may use a constitutive or seepage model to enforce physical structure while using ML to represent a difficult subcomponent, calibrate uncertain parameters or correct systematic residuals.

Validation and generalisation

A model should be tested on data that are independent in a meaningful sense. Randomly splitting neighbouring measurements can produce overly optimistic performance because spatially close data are correlated. Better tests may hold out an entire site, geological unit, project stage or time period.

Common metrics include mean absolute error, root mean square error and classification accuracy. For observations \(y_i\) and predictions \(\hat y_i\), the root mean square error is

\[

RMSE = \sqrt{\frac{1}{n}\sum_{i=1}^{n}(y_i-\hat y_i)^2},

\]

where \(n\) is the number of observations and the units are those of the predicted quantity. RMSE penalises large errors more strongly than mean absolute error. It does not, by itself, establish engineering suitability, unbiasedness or safe extrapolation.

Validation should also examine calibration, uncertainty intervals, sensitivity to input changes, performance for rare events and failure under distribution shift. A model trained on normally consolidated clay should not be assumed valid for structured, fissured or heavily overconsolidated clay merely because the input variables have similar numerical ranges.

Explainability and physics

Explainability concerns whether users can understand why a model produced a result and whether that reasoning is relevant to the engineering mechanism. Feature-importance scores can be useful but may be unstable or misleading when inputs are correlated. A visually simple explanation is not necessarily a causally correct explanation.

For safety-critical decisions, the engineer should be able to identify the data used, preprocessing steps, model version, applicable domain, uncertainty and reason for the recommendation. Physics-informed constraints, monotonicity where justified, dimensional consistency and conservation principles can improve credibility. They cannot guarantee correctness, but they reduce some forms of physically impossible prediction.

Uncertainty, risk and decision quality

Digital methods can reduce epistemic uncertainty by providing more observations, but they may introduce new uncertainty through algorithms, data transfer and model assumptions. Aleatory variability in soil properties remains even with perfect measurement.

A decision should consider both probability and consequence. A low-probability failure with catastrophic consequences may justify redundancy and conservative intervention. Conversely, excessive alarms can cause alert fatigue and lead operators to ignore genuine warnings.

Uncertainty should be communicated in a form appropriate to the decision: confidence intervals, prediction intervals, probability distributions, scenario ranges or qualitative confidence grades. A single predicted settlement value without uncertainty can create false precision. Where data are sparse, it is often more honest to report a range and identify the observations that would most reduce uncertainty.

Ethics, professional responsibility and governance

Professional accountability

Automation does not transfer responsibility from the engineer to software, a vendor or an algorithm. Professional duties include competence, protection of public safety, honesty about limitations, independent verification and clear communication of uncertainty. Engineers should not approve a model they cannot adequately assess, nor present a probabilistic forecast as a guarantee.

Bias and fairness

Bias can enter through sampling, historical project selection, missing data, sensor placement, labels, preprocessing or the decision rule itself. A model trained mainly on urban sites may perform poorly in rural or geologically different settings. Monitoring networks may overrepresent accessible locations and underrepresent vulnerable communities.

Bias assessment should compare performance across relevant geological, geographical and operational groups. If performance is inadequate for a group, the response may involve additional investigation, stratified modelling, conservative limits or refusal to automate the decision.

Data ownership and governance

A project should define who owns raw measurements, interpreted models, derived products and algorithm outputs. Governance should specify access rights, retention periods, version control, audit trails, correction procedures and rules for sharing sensitive information. Metadata should record source, units, coordinate reference system, datum, timing, processing, uncertainty and responsible party.

Personal, commercial and security-sensitive information requires appropriate protection. Geotechnical data may reveal critical infrastructure vulnerabilities, groundwater resources or construction sequencing. Open data can support research and reproducibility, but release should be balanced against privacy, security and contractual obligations.

Cybersecurity

Connected monitoring systems are part of critical infrastructure. Threats include unauthorised access, ransomware, sensor spoofing, denial of service and manipulation of alarms. Good practice includes least-privilege access, multi-factor authentication, encryption, network segmentation, secure software updates, backups, incident response plans and independent recovery testing. Safety functions should fail safely and should not depend exclusively on a remote cloud service.

Standards and regulation

Digital practice must be aligned with applicable national legislation, occupational safety requirements, data-protection law, procurement rules, professional codes and geotechnical design standards. Information-management standards, open exchange formats and quality-management systems can improve interoperability, but compliance with a standard does not prove that a particular model is fit for purpose. The responsible engineer must connect formal requirements with site-specific risk.

Sustainability and life-cycle value

Digitalisation can reduce unnecessary investigation, material use, rework, travel and unplanned maintenance. Continuous monitoring may enable performance-based intervention rather than premature replacement. However, sensors, servers, communications and repeated data processing consume energy and generate electronic waste. A life-cycle assessment should consider manufacture, installation, operation, maintenance, replacement and disposal.

Sustainable digital geotechnics therefore favours durable and repairable equipment, efficient sampling strategies, interoperable data, appropriate data retention and models that remain useful throughout the asset life. Digital tools should support climate adaptation by monitoring groundwater, settlement, erosion, permafrost degradation, extreme rainfall and changing slope conditions.

Illustrative case: automated monitoring of a deep urban excavation

Background

A fictional illustrative project involves a six-storey basement excavation adjacent to an existing masonry building and a major utility corridor. The site contains fill over compressible alluvial clay, with groundwater near the proposed excavation level.

Problem

Design calculations predicted manageable wall movement and settlement, but the consequences of excessive deformation were high. The project required timely detection of unexpected ground response without relying on a single instrument type.

Analysis and approach

The team established a baseline before excavation. The network included inclinometers behind the retaining wall, vibrating-wire piezometers in the clay, precise levelling points on the neighbouring building, robotic total-station prisms and rainfall records. Data were time-synchronised and transferred to a controlled database. A GIS displayed instrument locations and geological interpretation, while a BIM environment represented the excavation stages and structural sequence.

Predicted ranges were defined for wall displacement, pore pressure and building settlement at each excavation stage. An amber level required instrument verification and review of construction sequence; a red level required work suspension and implementation of a predefined stabilisation plan. Automated anomaly detection identified a sudden piezometer change, but the alert was not accepted immediately. The team checked the instrument against a nearby piezometer, reviewed recent pumping operations and inspected the cable and logger.

The apparent change was traced to a damaged cable connection rather than a true pore-pressure response. Later, a genuine increase in wall movement was detected by both an inclinometer and robotic total station. Construction was paused, strutting installation was brought forward and groundwater pumping was modified.

Outcome

The combined system prevented a false alarm from causing an unnecessary intervention and allowed a real deformation trend to be addressed before significant building damage occurred. The result depended on redundant observations, predefined actions and professional review, not on the anomaly algorithm alone.

Learning points

1. Data fusion is valuable only when measurements are independently credible.

2. Automated alarms require verification procedures and human oversight.

3. Trigger levels should be connected to construction decisions.

4. A digital representation must preserve assumptions, uncertainty and data provenance.

5. Cybersecurity, maintenance and communications continuity are safety issues.

Implementation framework for professional practice

A responsible digital-geotechnical programme can be developed through the following sequence:

1. Define the decision: Identify the hazard, decision-maker, intervention time and consequence of error.

2. Develop the ground model: Separate observations, interpretations and assumptions; identify uncertainty and failure mechanisms.

3. Select the minimum effective data: Choose instruments and surveys that address the decision rather than collecting data without purpose.

4. Specify quality requirements: Define calibration, accuracy, sampling, metadata, validation and maintenance.

5. Design the information architecture: Establish coordinate systems, naming conventions, access rights, version control and interoperability.

6. Test models independently: Use site-appropriate validation, uncertainty assessment and sensitivity analysis.

7. Define human roles: Assign responsibility for review, approval, intervention and communication.

8. Exercise failure scenarios: Test lost communications, false alarms, cyber incidents, sensor failure and unexpected ground behaviour.

9. Document and audit: Preserve data lineage, model versions, decisions and changes.

10. Review throughout the asset life: Update the model as new evidence becomes available and retire systems responsibly.

Emerging competencies and future directions

Future geotechnical professionals will need stronger competence in sensor technology, statistics, coding, databases, GIS, BIM, remote sensing, model verification and cybersecurity. Equally important are communication, systems thinking, ethics and the ability to recognise when a model is outside its domain.

Emerging developments include edge computing for local decisions, fibre-optic distributed sensing, autonomous site inspection, probabilistic digital twins, physics-informed learning, robotic investigation, large-scale satellite monitoring and standardised data exchange. These developments are likely to increase productivity, but they also intensify the need for governance and independent assurance.

The most valuable professional remains not the person who uses the most advanced tool, but the person who can integrate evidence, mechanics, uncertainty and consequences into a defensible decision. Future practice will therefore be hybrid: field observation and computation, physical models and data-driven models, automation and accountable judgement.

Chapter Summary

Digital geotechnics links sensing, data management, modelling and decision-making across the life cycle of geotechnical assets. Sensor networks and remote sensing improve temporal and spatial observation, but require calibration, redundancy, metadata and quality control. GIS supports spatial reasoning; BIM supports coordinated information management; and digital twins connect physical assets with continuously updated models and decisions.

Machine learning can assist classification, prediction, anomaly detection and parameter estimation, particularly when combined with mechanics. Its performance must be tested for generalisation, uncertainty, bias and distribution shift. Explainability is essential for safety-critical use, but no algorithm removes professional responsibility.

Ethical digital practice requires accountability, fairness, data governance, cybersecurity, compliance with applicable standards and transparent communication of uncertainty. Sustainability requires consideration of the complete life cycle of digital equipment and data. The future of geotechnical engineering will depend on professionals who can combine soil mechanics with data literacy, systems engineering and ethical judgement.

Key Terms

Digital geotechnics: Use of connected digital data, models and systems in geotechnical engineering.

Sensor network: Interconnected instruments that measure physical or environmental variables.

Data provenance: Record of the origin, processing and history of a dataset.

Digital thread: Traceable link between observations, models, decisions and project stages.

GIS: System for storing, analysing and visualising spatially referenced information.

BIM: Structured information environment for coordinated design, construction and asset management.

Digital twin: Dynamic digital representation linked to a physical asset and its observations.

Remote sensing: Measurement of Earth or infrastructure properties without direct contact.

InSAR: Radar-interferometry technique used to estimate surface displacement.

Machine learning: Computational methods that infer patterns or relationships from data.

Supervised learning: Learning from examples with known input-output relationships.

Distribution shift: Change between the data used for training and the conditions of application.

Explainability: Ability to understand and communicate how a model reaches a result.

Model uncertainty: Uncertainty arising from simplifications or inadequacy of a model.

Measurement bias: Systematic difference between a recorded and true quantity.

Observational method: Design approach based on prediction, monitoring and predefined response.

Data governance: Rules for ownership, access, quality, security, retention and use of data.

Alert fatigue: Reduced attention caused by excessive or unreliable warnings.

Physics-informed model: Model constrained or guided by physical principles.

Interoperability: Capacity of systems to exchange and correctly interpret information.

Review Questions

Short Answer Questions

1. Why is sensor redundancy important in a critical geotechnical monitoring system?

2. Distinguish between BIM and a digital twin.

3. State two limitations of InSAR for geotechnical monitoring.

4. What is meant by distribution shift in machine learning?

5. Identify four essential elements of geotechnical data governance.

Long Answer Questions

1. Explain how an automated monitoring system should be designed from failure mechanisms and intervention decisions.

2. Critically compare GIS, BIM and digital twins in the management of geotechnical information.

3. Discuss the opportunities and limitations of machine learning for soil-parameter prediction.

4. Explain how explainability, uncertainty and professional accountability should influence the use of AI in geotechnical engineering.

5. Develop a cybersecurity and quality-assurance strategy for a connected monitoring system at a major excavation.

Analytical / Application Questions

1. A retaining-wall monitoring system reports a sudden displacement increase from one inclinometer, while adjacent instruments remain stable. Propose a technically defensible investigation and decision sequence.

2. Design a monitoring and data-governance framework for a landslide-prone transport corridor using ground sensors, InSAR, GIS and machine learning.

3. A settlement-prediction model performs well on its training projects but poorly on a new site with organic clay. Analyse the likely causes and recommend a validation and risk-control strategy.

Multiple Choice Questions with Answers

Question 1

Which feature most clearly distinguishes a digital twin from a static BIM model?

A. Use of three-dimensional geometry

B. Connection to physical observations and updating processes

C. Storage of construction drawings

D. Representation of structural elements

Correct Answer: B. Connection to physical observations and updating processes

Question 2

What is the principal meaning of the bias term in the measurement model y = x + b + ε?

A. Random sampling variability

B. A systematic measurement offset

C. The physical quantity being measured

D. The number of observations

Correct Answer: B. A systematic measurement offset

Question 3

Which validation approach best tests whether a geotechnical ML model generalises to a new site?

A. Randomly duplicate observations from the training site

B. Test only on the data used for training

C. Hold out data from an independent site

D. Increase the number of model parameters

Correct Answer: C. Hold out data from an independent site

Question 4

InSAR displacement is normally measured directly along which direction?

A. The vertical direction only

B. The horizontal north direction only

C. The radar line of sight

D. The direction of groundwater flow

Correct Answer: C. The radar line of sight

Question 5

What is the main purpose of a green-amber-red monitoring scheme?

A. To eliminate all measurement uncertainty

B. To link observations with predefined responses

C. To replace engineering judgement

D. To increase the number of sensors automatically

Correct Answer: B. To link observations with predefined responses

Question 6

For predictions and observations, RMSE is calculated by taking the square root of the:

A. Mean of absolute errors

B. Mean of squared errors

C. Sum of signed errors

D. Median of squared observations

Correct Answer: B. Mean of squared errors

Question 7

Which practice most directly reduces the risk of unauthorised access to monitoring data?

A. Using identical passwords for all users

B. Disabling software updates

C. Applying least-privilege access and multi-factor authentication

D. Publishing all system credentials

Correct Answer: C. Applying least-privilege access and multi-factor authentication

Question 8

Why can random splitting of neighbouring geotechnical data produce misleading ML performance?

A. Neighbouring observations may be spatially correlated

B. Random splitting always removes all data

C. It prevents any model from being trained

D. It changes soil grain size

Correct Answer: A. Neighbouring observations may be spatially correlated

Question 9

Which statement about remote sensing is correct?

A. It always replaces ground investigation

B. It provides measurements unaffected by atmosphere or vegetation

C. Its results should be interpreted with regard to geometry and control

D. It directly measures all three displacement components

Correct Answer: C. Its results should be interpreted with regard to geometry and control

Question 10

Who retains professional responsibility when an engineer uses an AI-based geotechnical recommendation?

A. Only the software vendor

B. Only the data-storage provider

C. The responsible professional and project organisation

D. No party, if the algorithm is accurate on average

Correct Answer: C. The responsible professional and project organisation