Data leakage is the most common cause of results that look excellent and fail in use. In reservoir applications it takes specific forms: random splitting of production time series so that adjacent months appear in both training and test sets, splitting by well when wells from the same pad share completion and geology, and including features that encode the answer such as cumulative production when predicting ultimate recovery. Models validated with leakage report accuracy that has no relationship to their performance on genuinely new data, and the reported accuracy is often what secures further investment in the approach.

Sample size is the binding constraint in most subsurface applications. A field with two hundred wells provides two hundred examples, and many of them are correlated because they share geology, completion design and operating practice. Flexible methods with many parameters will fit that data set perfectly and generalise poorly. Simpler methods with fewer parameters, or methods that incorporate physical constraints, generally perform better, and the instinct to reach for the most capable available algorithm is usually wrong here.

Physics is available and should be used. Reservoir problems are governed by conservation laws, flow equations and equipment characteristics that are known, and a purely data-driven model that ignores them is discarding information. Hybrid approaches that use a physical model to establish expected behaviour and machine learning to correct its residual, or that constrain a learned model to respect physical bounds, consistently outperform purely data-driven approaches on subsurface problems and extrapolate far better.

Finally, interpretability is not a preference in this field. A model recommending a well location, a completion design or a reserves booking will be reviewed by engineers, partners and auditors who will ask why it produced that answer. A model that cannot explain itself will not be used for consequential decisions regardless of its accuracy, which makes interpretable methods and explanation techniques a practical requirement rather than an academic concern.

By the end of this training, participants will be able to:

• Prepare subsurface data for machine learning including missing data, outliers, scaling and feature construction
• Identify and prevent data leakage in spatially and temporally correlated reservoir data
• Select supervised methods appropriate to sample size, correlation structure and interpretability requirement
• Apply unsupervised methods for clustering, dimensionality reduction and anomaly detection
• Design validation schemes appropriate to correlated subsurface data
• Diagnose overfitting and assess genuine generalisation performance
• Build physics-informed and hybrid models that combine engineering knowledge with learned components
• Construct machine learning proxies for reservoir simulation and validate their range of validity
• Apply interpretability methods and explain model behaviour to reviewers
• Deploy and govern models used in reservoir engineering decisions

The training is worked on subsurface data sets throughout, using well, production and completion data rather than generic examples. Participants build models, and the same data sets are then re-analysed with correct validation schemes to demonstrate how much of the apparent performance was leakage or overfitting. Physics-informed and purely data-driven approaches are applied to the same problems and compared on extrapolation. Published subsurface machine learning applications are examined critically for validation practice. Interpretability methods are applied to models participants have built.

Organisations sending participants to this training will:

• Apply machine learning to subsurface problems where it genuinely adds value
• Avoid investment in approaches whose reported performance came from validation error
• Improve forecasting and screening workflows through appropriate hybrid methods
• Reduce simulation cost through validated proxy models
• Build internal capability rather than depending on external providers for subsurface analytics
• Establish governance so that models used in decisions can be reviewed and defended

Participants will:

• Build machine learning models on subsurface data correctly
• Recognise leakage, overfitting and inflated performance claims
• Combine physical understanding with learning methods
• Validate models in a way that reflects real predictive performance
• Explain model behaviour to engineers and reviewers
• Build a capability at the intersection of reservoir engineering and data analysis

• Reservoir and production engineers working with field data
• Simulation engineers building proxies and uncertainty workflows
• Data scientists working on subsurface problems
• Geoscientists applying quantitative methods to reservoir data
• Technical staff evaluating machine learning proposals and vendor claims
• Digital and analytics staff supporting subsurface teams
• Engineers with limited programming background entering data analysis work

Upon successful completion of this training course, delegates will be awarded an official Certificate of Completion issued by the Petroleum Engineers Association (PEA), an ISO 9001:2015 certified training organization. The certificate carries 10 Credits and formally records the total learning hours completed.

Each certificate is signed by the Course Facilitator and the CEO of the Petroleum Engineers Association, and serves as verifiable proof of professional training that delegates can present to employers and professional bodies worldwide.