Machine Learning Beyond the Lab: Real Uses in Physics, Service, Earth Science, and Medicine

Machine Learning Stopped Being a Niche Topic

For a long time, machine learning sat in a fairly small corner of computer science. Researchers trained models. Students ran homework experiments. A few companies used recommendation engines. That picture is outdated. The interesting story now is not that machine learning exists. It is that the same family of techniques is showing up in physics labs, customer support queues, satellite image analysis, drug screening, and connected devices in homes and factories.

Kais Othman and Maroua Ben Yahia, writing from the Higher Institute of Computer Science at the University of Tunis El Manar, make that spread the subject of a 2021 review in the World Journal of Advanced Engineering Technology and Sciences. Their research article does not try to invent a new algorithm. It maps how machine learning is already being used, which tools people actually reach for, and which problems keep coming back: privacy, bias, and decisions that nobody can explain after the fact.

That mapping still matters. Teams often talk about “doing AI” as if it were one product. It is not. A symbolic regression method that recovers Newton’s second law from data is a different job from a chatbot that answers billing questions. A model that flags likely mineral deposits is a different job from a network that predicts whether a natural compound will be toxic. The shared ingredient is pattern finding at a scale people cannot do by hand. The settings, risks, and payoffs are not the same.

"The same family of methods can recover a physical law, route a support ticket, or screen a molecule. The method is shared. The stakes are not."

This piece walks through the review in plain language. You will see how machine learning is used in physics, customer service, geosciences, smart systems, and drug discovery. You will also see why TensorFlow and PyTorch keep coming up, and why ethics is not an appendix. If you build software, manage a product, or just want a clearer picture of where these models actually sit in the world, this is a useful tour.

Why a Cross-Field Review Is Useful

Most machine learning writing stays inside one industry. Healthcare papers talk to healthcare readers. Marketing papers talk to marketers. That is fine for specialists. It is a weak way to notice what is actually transferring. When you put physics next to customer service next to geology, the common moves become easier to see. So do the common failures.

Othman and Ben Yahia treat machine learning as a general-purpose way to process large datasets, find structure, and make predictions. That sounds obvious. In practice it is the reason the same toolkit can move from one domain to another. Physics needs models that respect units and conservation laws. Support teams need systems that stay available at 2 a.m. Earth scientists need models that can handle space and time together. Drug researchers need predictions early enough to avoid expensive late-stage failures. Different constraints. Same hunger for patterns in messy data.

The review also puts software engineering in the picture. Machine learning is not only a research method. It is now part of how software is built, tested, deployed, and maintained. That is an easy thing to miss if you only read application papers. Once models sit inside products, someone has to choose a framework, watch data quality, and decide what happens when the model is wrong. Those are engineering questions, not just science questions.

One more reason the review is worth reading: it refuses the “everything is solved” tone. The authors keep returning to privacy, biased training data, and automated decisions that are hard to inspect. That is the right instinct. A method that works in a paper can still be a bad idea in a hiring tool, a loan system, or a medical workflow. Capability and permission are not the same thing.

Physics: Teaching Machines to Recover Equations

Physics is a surprising place to start, and that is the point. People often assume machine learning is for images, ads, and language. The review opens with something more old-fashioned: finding the equation that explains a dataset. That task is called symbolic regression. Instead of outputting a probability or a label, the model tries to output a formula.

The headline example is AI Feynman, a method from Udrescu and Tegmark. It is not a generic black box dropped onto physics data. It is built around habits physicists already use. First, dimensional analysis. If an equation mixes meters with seconds in a way that cannot be right, throw it out. That single filter cuts a huge search space. Second, separability. If a complicated expression can be split into smaller pieces, solve the pieces. Third, brute-force search, but only after the smarter filters have done their work. Exhaustive search is expensive. Using it last is the whole trick.

Why does that matter? Because many scientific datasets are not waiting for a classifier. They are waiting for a relationship. If a system can recover Newton’s second law or conservation of momentum from raw numbers, it is doing something different from “predict the next value.” It is proposing a compact description of the world. That description can be checked, argued with, and reused. A neural net that merely interpolates well is harder to take into a theory seminar.

The practical angle is easy to miss. Sensor streams in engineering look a lot like experimental physics data. If a method can extract a governing relationship, it can also support predictive maintenance, control systems, and simulation. You do not need a new law of nature for the method to be useful. You need a compact model that a person can inspect. That is often more valuable than a slightly better black-box score.

Looking ahead, the review expects these methods to face messier datasets: quantum systems, cosmology, noisy experiments. That is the right next test. Recovering a textbook law is impressive. Recovering a useful approximation in a setting where no clean law is waiting would be more useful still. The honest caveat is that symbolic regression is still hard. Many relationships are not neat. Many datasets are incomplete. A method that works on well-posed physics problems will not automatically work on every spreadsheet.

Customer Service: Chatbots That Change How People Comply

From equations to help desks. The jump is large, and that is useful. In customer service, the model is not looking for a law of nature. It is looking for an answer a person will accept, quickly, at any hour. Othman and Ben Yahia lean on work by Adam, Wessel, and Benlian on AI chatbots and user compliance. The interesting word there is compliance. The claim is not only that bots are cheaper. It is that they can change whether people follow the process in front of them.

The mechanics are familiar. Natural language processing reads the request. A model maps it to an intent. The system replies in a conversational style. The operational gains are also familiar: 24/7 coverage, many conversations at once, consistent wording, fewer repeated answers from humans. The review is clear about the division of labor. Bots take the routine load. People take the messy, sensitive, or high-stakes cases. That split only works if the handoff is clean. A bot that traps a frustrated customer in a loop is not “automated service.” It is a new kind of wait time.

Personalization is the second half of the story. If the system can see prior purchases or browsing history, it can skip the generic script. That can feel helpful. It can also feel invasive. The same data that makes a reply feel tailored is the data that makes privacy a live issue. The review does not pretend those two facts cancel out. Speed and personalization raise satisfaction when they work. They raise risk when the data is wrong, stale, or used in a way the customer never agreed to.

There is a mild contradiction here, and it is worth keeping. Chatbots can make service feel more personal while also making it more standardized. The reply is tailored to you, but the tailoring is a template. For many tasks that is enough. For a billing dispute after a death in the family, it is not. The review’s future view is that better language models will take on more complex work. That may be true. It does not remove the need for an escape hatch to a person who can actually decide.

If you run a support team, the practical lesson is not “add a bot.” It is “decide which intents are safe to automate, measure containment without hiding failed handoffs, and keep a human path that does not feel like a punishment.” The research is about technology. The operating system around it is still a service design problem.

Geosciences: Reading the Planet at Scale

Earth science has a data problem that looks a lot like an opportunity. Satellites, drones, and ground sensors produce more measurements than a person can inspect. The measurements also have two awkward properties: they sit in space, and they change over time. A field that used to depend on expert interpretation of maps now depends on models that can classify land cover, watch vegetation change, estimate water quality, and help forecast hazards.

The review points to standard tools in this setting: neural networks, decision trees, and support vector machines. The names are less important than the job. These models find patterns that are easy to miss in high-dimensional data. They can combine temporal and spatial signals into one prediction. That is hard to do with a single static map. It is also hard to do with a person scrolling through years of imagery.

The case studies are concrete. Rainfall and storm models trained on historical weather data can improve preparedness. Seismic models can estimate aftershock risk after a main shock. Mineral exploration models can rank likely locations and cut the cost of looking in the wrong place. None of those examples is science fiction. They are the kind of prediction work geosciences already do, sped up and, in some cases, made more accurate.

There is a catch, and the review names it. Data is not always accessible. Sources do not always line up. Environmental systems change. A model trained on last decade’s climate is not automatically valid for the next one. That is not a reason to avoid machine learning. It is a reason to treat model refresh, sensor calibration, and uncertainty as first-class work. A confident map that is wrong can be worse than a slower expert judgment, especially when the map is used for disaster response or resource allocation.

The longer view is real-time analysis. As sensors get better and compute gets cheaper, the lag between “something happened on the ground” and “someone can act on it” should shrink. That is the promise. The work that makes the promise real is unglamorous: data pipelines, quality checks, and models that degrade honestly instead of failing silently.

Smart Systems: IoT, 5G, and Decisions at the Edge

Smart systems are where machine learning stops looking like a research method and starts looking like infrastructure. The review focuses on the Internet of Things and 5G. The pairing is not accidental. IoT produces streams of sensor data. 5G makes it easier to move that data quickly. Machine learning is the layer that turns the stream into an action: dim the lights, reroute traffic, flag a failing motor, alert a clinician.

A smart home is the friendly example. Models learn when people are home, how they use heat and light, and what “normal” looks like for a security sensor. The point is not novelty. The point is a system that adapts without a person rewriting rules every week. Industrial IoT is the harder example. Predictive maintenance can keep a factory from stopping because one machine failed without warning. That is money, safety, and schedule all at once.

5G changes the timing. Some applications cannot wait for a round trip to a distant cloud. Autonomous vehicles and remote care are the usual illustrations. If a decision has to happen now, the model has to be close to the device, or the network has to be fast and reliable enough that “close” is less important. The review, drawing on Whitmore, Agarwal, and Da Xu, treats this mix of IoT and 5G as the path toward smarter cities and more responsive public services. Traffic, utilities, and emergency systems are all candidates. So is network management itself: allocating bandwidth and spotting threats in traffic that is too large to inspect by hand.

This is the section where “more connected” can quietly become “more fragile.” A model that saves energy in a building is a win. A model that locks people out because a sensor misfires is not. The review’s useful pressure is to keep asking whether the system can operate under messy conditions, not only in a demo. Robustness and energy use are not side quests. They decide whether the smart system is actually smarter than the dumb one it replaced.

Drug Discovery: Faster Paths Through Natural Compounds

Drug discovery is slow, expensive, and full of late surprises. Natural products make that worse and better at the same time. They are a rich source of new chemistry. They are also structurally complex and hard to get in useful amounts. Machine learning enters as a filter. Instead of testing everything in a wet lab first, models look for likely activity, likely failure, and likely toxicity earlier in the process.

A key target is ADMET: absorption, distribution, metabolism, excretion, and toxicity. If a compound cannot be absorbed, or is likely to be toxic, it should not eat years of budget. The review argues that machine learning can cut late-stage failure by making those calls earlier. That is not a claim that models replace chemists. It is a claim that models can rank options so chemists spend time on better candidates.

The examples, drawing on Newman and Cragg’s long view of natural products as drug sources, are specific. Deep learning can search chemical libraries for inhibitors tied to cancer or infectious disease, then suggest structural tweaks. Other models help unpack biosynthetic pathways so researchers can engineer microbes to make compounds that are scarce in nature. That second use is easy to overlook. Discovery is not only “find a molecule.” It is also “make enough of it to study.”

The next wave named in the review includes reinforcement learning and generative models. Those methods do not only score existing compounds. They can propose new ones. That is exciting and easy to oversell. A generated structure is not a drug. It is a hypothesis that still has to survive synthesis, assays, safety work, and regulation. The review is at its best when it treats machine learning as an accelerator inside a biological process, not as a replacement for that process.

Interdisciplinary work is not a slogan here. It is the requirement. A model that is statistically elegant and biologically nonsense is a waste. The authors stress collaboration between computational scientists and biologists so that predictions stay actionable. If you work adjacent to this field, the takeaway is simple. The bottleneck is shifting from “we cannot look at that many compounds” to “we cannot validate that many promising hits.” Tools that create more candidates without a validation plan just move the pile.

The Tooling Layer: TensorFlow, PyTorch, and the Rest

Applications get the headlines. Frameworks decide whether the work ships. The review’s survey of tools, based in part on Nguyen and colleagues, is one of the more practical sections. TensorFlow and PyTorch dominate the discussion, and for understandable reasons. They have communities, hardware support, and enough examples that a team can hire people who already know them.

TensorFlow, from Google, is built around computational graphs and a path from experiment to production. If you need to run the same model on different hardware, and then keep it alive in a product, that design is attractive. PyTorch, from Facebook at the time of the paper, builds graphs dynamically. That makes debugging and prototyping feel closer to ordinary Python. Researchers often prefer it for that reason. The split is not a morality play. It is a workflow choice. Rapid iteration in a lab is not the same job as a stable service in production.

The review does not pretend those two libraries are the whole market. Microsoft’s CNTK and Apache MXNet appear as other options with their own strengths. The point is not to memorize a catalog. The point is that the “right” framework depends on the team, the deployment target, and the surrounding software. A brilliant model in a framework nobody on the team can maintain is not a brilliant product.

There is also a quieter claim: machine learning is moving into the software lifecycle itself. Code generation, testing, and maintenance are all candidates for assistance. That idea sits next to the rest of the paper rather than inside one industry vignette. If you only remember one tooling lesson, remember this. Choosing a library is part of system design. It affects hiring, monitoring, and how painful the second year of the project will be.

The Hard Part: Privacy, Bias, and Decisions Nobody Can Explain

Every section above gets more interesting, and more dangerous, when you ask who is affected. The review’s ethics chapter is not decorative. It names three problems that show up in almost every domain: data privacy, model bias, and automated decisions that lack a clear account of why they happened.

Privacy is the first. Training needs data. A lot of the most useful data is personal or commercially sensitive. Techniques such as differential privacy and federated learning exist because of that tension. They are not magic. They can protect information and still cost accuracy. Teams that ignore the tradeoff usually discover it later, in a breach report or a regulator’s letter.

Bias is the second. Models learn from history. If history is skewed, the model can repeat the skew with more confidence and more speed. Hiring, lending, and policing are the usual warning cases, and they deserve the warning. The same issue appears in quieter places. A geoscience model trained on well-instrumented regions can under-serve places with fewer sensors. A chatbot trained on one dialect can fail another. Bias is not only a social-media controversy. It is a coverage problem in the training set.

Explainability is the third. Healthcare, criminal justice, and finance already use models to rank, approve, or flag. When the stakes are high, “the network said so” is not an answer. The review calls for systems that can produce an account a person can inspect. That is still an open research problem. It is also a product requirement. If a clinician or a loan officer cannot challenge the output, the system is not a decision aid. It is a hidden policy.

Regulation sits over all three. GDPR is the example the authors use, because it forces a conversation about lawful use, access, and purpose. Compliance is not the same thing as ethics, but it is the floor many organizations will actually meet. The review also flags employment. Automation that removes routine work can raise productivity and still concentrate harm on the people who used to do that work. “The model is accurate” does not settle that question.

The recommended response is multidisciplinary on purpose: ethicists, lawyers, engineers, and policymakers in the same room. That can sound like a committee. The alternative is a model that ships because it can, not because it should. Education matters too. If the public only hears slogans, the debate stays stuck between hype and panic. Clearer explanations of what these systems do, and where they fail, are part of the work.

What This Adds Up To

Read the five application areas together and a pattern shows up. Machine learning is at its best when the dataset is large, the pattern is real, and a person still owns the decision that follows. Physics wants inspectable formulas. Support wants fast answers with an escape hatch. Geosciences want predictions that can be checked against the next storm or aftershock. Smart systems want actions that still make sense when a sensor is dirty. Drug discovery wants fewer bad bets, not a fantasy of drugs designed in a single click.

Othman and Ben Yahia’s contribution is the map, not a single breakthrough. In a field that often sells the next architecture, a careful tour of where the current tools already sit is more useful than it sounds. You can read the same research article as a snapshot of 2021 and still use it as a checklist in 2026: where is the data, which framework will you live with, what happens when the model is wrong, and who is accountable for the output.

The future the authors sketch is easy to agree with: better models, more data, more domains, more collaboration. The useful caution is also theirs. Wider use will not automatically be fairer or safer. It will only be wider. If you take one idea from this review into your own work, take that. Machine learning can travel across fields. Responsibility has to travel with it, or the same method that recovers a physical law can just as easily scale a bad decision.

That is the real story here. Not that machines can learn. We already knew that. The story is that learning systems are now part of how science, service, infrastructure, and medicine get done, and that the people who ship them still have to decide what “done” is allowed to mean.

Frequently Asked Questions

What fields does this machine learning review cover?

Othman and Ben Yahia review machine learning in physics, customer service, geosciences, drug discovery, and smart systems that use IoT and 5G. They also compare common frameworks such as TensorFlow and PyTorch and discuss privacy, bias, and automated decisions.

What is AI Feynman in physics?

AI Feynman is a physics-inspired symbolic regression method. It uses dimensional analysis, separability, and targeted search to recover formulas from data, including known laws such as Newton’s second law.

Which machine learning frameworks does the paper compare?

The review focuses on TensorFlow for scalable production graphs and PyTorch for dynamic graphs used in research prototyping. It also notes CNTK and MXNet as additional options.

What ethical issues does the paper raise?

The authors highlight data privacy, biased training data, and opaque automated decisions. They also point to GDPR-style rules and the employment effects of automation.

Related articles

• Advances in machine learning and software engineering
• AI-based modeling in software engineering
• Data-driven decisions with IoT and AI