Editor's Note: Abdulmalik Ajibade is a member of the TWA Editorial Board and a contributing author of previous TWA articles.
Domain experts routinely detect patterns that may not appear on dashboards. They recognize abnormal production behavior, identify inconsistencies in subsurface interpretations, and understand which variables are operationally meaningful long before a model is trained. Yet transforming these insights into machine-learning (ML) solutions often remains inaccessible due to programming requirements, tooling complexity, and data engineering overheads.
No-code ML platforms have fundamentally challenged the status quo and changed the pervasive narrative. Rather than lowering standards, they shift the entry point of ML from software proficiency to domain understanding. These tools enable subject-matter experts across engineering and geoscience disciplines to prototype, evaluate, and stress-test ML concepts without writing any code. This allows them to focus more on the problem rather than the process.
This article examines how domain experts can use no-code ML platforms to explore decision-relevant problems, validate hypotheses, quickly build prototypes, and engage more effectively with data science teams when solutions transition toward production.
What Are No-Code Machine Learning Platforms?
No-code ML platforms provide visual environments for constructing predictive modeling workflows using predefined components for data ingestion, preparation, and analysis on one hand and model training, selection, validation, evaluation, and deployment on the other. Users configure models through structured interfaces rather than custom code.
A distinction is often made between no-code and low-code platforms. No-code systems rely entirely on visual configuration, while low-code platforms allow optional scripting (another word for coding) for advanced customization. Both significantly reduce the technical barrier relative to traditional ML-development pipelines.
These platforms are not substitutes for sound methodologies. They do not resolve data-quality issues, define appropriate problem formulations, or eliminate the need for validation. Their value lies in accelerating early-stage experimentation, feasibility assessment, and prototyping, not bypassing foundational principles.
How No-Code ML Works in Practice
Despite vendor differences, most no-code platforms follow a consistent workflow. Firstly, the relevant data for the identified problem or challenge are identified and located. This is data selection. This is fundamental to the entire process. This is critical to ensuring that the right physics is ingested into the solution space of the problem, the right premise is provided to answer the research question, and the right input goes in to produce a reliable output. Then the identified data are ingested from operational or historical sources such as time-series measurements, logs, or structured records. Visual exploration tools enable users to assess the data distributions, identify missing values, and evaluate basic data integrity indices. This is data preparation and analysis.
Feature selection and preparation follow. While automated options are available, domain judgment remains essential and critical for this step. Decisions regarding variable relevance, anomaly handling, aggregation windows, and transformations require contextual understanding that cannot be fully and reliably automated. Model training typically relies on AutoML techniques, where multiple algorithms are evaluated and hyperparameters tuned automatically. Domain experts define the study objectives, prediction targets, logical and contextual constraints, validation priorities, and evaluation metrics. Trade-offs between false positives and false negatives, and interpretability and stability must be explicitly considered.
Meticulous evaluation of the models’ performance allows the identification of the best model with respect to the problem being addressed. This is model selection. It is the standard ML-modeling practice not to rely on a single algorithm for all problems or data as different algorithms will exhibit different performances with different data. It should be noted that there is no universally super or best algorithm in all data conditions. For each data or problem, the most optimal model must be investigated and identified. This is called the "no-free-lunch theorem" in the ML community (Sterkenburg and Grunwald, 2021).
Model validation closes the loop. Performance metrics, prediction diagnostics, and feature-importance analyses help determine whether the selected model’s behavior is both statistically credible and physically plausible in meeting the set objectives of the study or addressing the identified challenge. The outcome of this critical step decides whether the model’s performance is acceptable and can be launched into production for operational use or if there is the need to revisit any of the previous steps to improve the model’s performance or investigate another possibly best model.
These steps are summarized in Fig. 1.
Practical Principles for Domain Experts
Effective use of no-code ML requires discipline and caution. Firstly, problems should be narrowly scoped and objective-driven. Early success depends on addressing clearly-defined questions with measurable outcomes. Secondly, evaluation metrics must reflect operational relevance and realities. High aggregate accuracy can obscure failure modes that matter most in practice. Third, results should be interrogated through domain logic. Models that contradict physical understanding or operational constraints require reassessment, regardless of numerical performance. Finally, the development process should be documented rigorously. Transparency is essential when transitioning prototypes to data science or engineering teams for production deployment.
The example application discussed in the next section summarizes these principles.
Example: No-Code ML Platform Implementation for Subsurface Applications
We showcase a practical implementation of a no-code platform, AiLAB (Anifowose et al., 2025). This platform allows users to
- Optionally connect to the corporate database to fetch data (such as formation tops) or upload them from local or shared folders.
- Visualize data relationships.
- Clean and analyze data.
- Build and train several ML models (regression, classification, and clustering).
- Evaluate the models and select the best, based on performance metrics.
- Save the best model and use to it in production mode to perform further blind validations or make new predictions.
In this section, we present a quick example of the use of this AiLAB platform to investigate the feasibility of predicting total porosity (PHIT) (target) from advanced mud gas data (input features). While the ground truth for most reservoir rock or fluid properties is measured directly on the respective samples obtained from specific intervals of interest and analyzed through extensive laboratory procedures, the measurements are limited and sparse, and the analysis are typically costly, time-consuming, and labor-intensive. Before the emergence and wide application of ML methodologies, empirical and statistical relationships have been used to achieve the same objective. However, these latter approaches have limitations. Some of the empirical equations were derived based on certain assumptions on the data distribution or linearity in the relationship between the input features and the target to ensure simplicity and explainability. Some of the assumptions may not be applicable or obtainable in real operational scenarios or geological settings in all situations. These would introduce bias into the modeling process.
Starting with the visualization feature, Figs. 2 and 3 exemplify and demonstrate the core principles discussed in the last section. Domain experts should ensure that the selected data and then features correlate with the target attributed to be modeled. Fig. 2 shows that some of the input variables/features possess enough relevance and correlation to answer the question of whether it is possible to predict PHIT from advanced mud gas data (AMG), including the light and heavy components.
As part of understanding the data, the AMG data is a product of the mud logging process. It includes identifying and quantifying the individual natural gas components dissolved in the mud while drilling oil and gas wells and circulated to the surface during hole cleaning. While the advanced mud gas data includes the heavier components (C6+) such as heptane, octane, benzene, toluene, and methylcyclohexane, the basic gas data are limited to the C1 (methane) to C5 (pentane).
Traditionally, the basic and advanced mud gas data have been used extensively for reservoir fluid characterization. This example helped to confirm that their utilization can be extended beyond the latter to predict reservoir rock properties, such as porosity, especially in real time, since the AMG data are also obtained in real time.
Fig. 3 demonstrates the model validation and selection processes, where relevant numerical (such as correlation coefficient, differential error, etc.) and qualitative metrics (such as crossplots) are employed to determine the best model for further validation and selection for prediction. In this example, and in proper consideration of the no-free-lunch theorem, three ML algorithms, including Feedforward Neural Network (FForwardNN) (Fig. 3(b)), regression tree Reg Tree) (Fig. 3(c)), and random forest (RanFor) (Fig. 3(d)) were implemented in comparison with a linear regression (LinR) (Fig. 3(a)) algorithm. This comparison with linear regression is meant to confirm that the relationship between the gas components and PHIT is so complex that a linear regression model would not be capable of handling it. The visual comparison of the crossplots confirms that the RanFor model performs the best in terms of training and testing/validation. Hence, this model could be used for further validation with other wells or to make predictions in a new well being drilled.
Conclusion
No-code ML platforms offer domain experts, as well as young professionals, an accessible way to work directly with the machine learning methodology without the hassles and challenges associated with coding. They do not replace data scientists, nor do they eliminate the need for robust engineering when solutions scale from prototypes to production modes. Their contribution lies in improving problem formulation, accelerating validation, and ensuring that ML efforts originate from operational reality rather than abstract modeling.
The primary constraint is no longer access to tools, but clarity of purpose. Machine-learning value emerges not from model complexity, but from alignment with decisions, workflows, and constraints.
For Further Reading
AiLAB: An In-House Developed Codeless Platform for Machine Learning Modeling by F. Anifowose, S. Badawood, and M. Mezghani.
The No-Free-Lunch Theorems of Supervised Learning by T. Sterkenburg and P. Grunwald.
Disclaimer: The views expressed in this article are those of the authors and in no way represent those of their affiliates.
Abdulmalik Ajibade is a reservoir engineer at Beicip-Franlab, France, specializing in digital transformation and AI applications for subsurface and production assets. Ajibade is a certified data analyst and volunteers as a machine-learning engineer at Omdena. He was the winner of the SPE Nigeria Paper Contest for his work on improving production optimization with machine learning. His work focuses on integrating data-driven and AI workflows to enhance asset performance and decision-making in the energy sector. He holds an MSc in geo-data management for the energy mix from IFP School, France, and a BSc in petroleum and gas engineering from Nile University of Nigeria.
Fatai Anifowose, SPE, is an independent researcher in the field of machine learning. His research focuses on automating geological and petroleum engineering workflows and application of machine learning to increase accuracy, improve efficiency, and enhance productivity. His accomplishments include over 90 papers, more than 10 granted patents and several filed, and a number of R&D awards, including the 2021 SPE Middle East and North Africa Regional Service Award, 2021 SPE Middle East and North Africa Regional Data Science and Engineering Analytics, and 2024 IChemE Learning and Development Award. He is a technical reviewer for international conferences and journals. He is a member of EAGE, SPE, AAPG, and Dhahran Geoscience Society.