MENU
Advanced Materials Intelligence
  1. Home
  2. Read This First: AI and machine learning — An Introduction

Read This First: AI and machine learning — An Introduction

← Back to this field’s guide library

TROY TECHNICAL · BEGINNER’S GUIDE

READ THIS FIRSTAI and machine learning — An Introduction

What does AI learn, and what does it not know? A first guide beyond the chat window.

For newcomers / The big picture before equations / Created 17 October 2026

Japanese version →

Finding defects in photographs, predicting a material property and drafting text look like different activities. Machine-learning applications share a useful starting idea: examples are used to build a way of responding to inputs not yet seen. Before equations or product names, this guide connects that mechanism to the decisions a user must make.

01Learning a relationship instead of writing every rule

Suppose the task is to predict strength from a material formulation. A conventional program may apply an equation or explicit rules chosen by an engineer. A machine-learning system adjusts internal numerical parameters using examples that pair formulations with measured strengths. That adjustment is training. The resulting mechanism is a model; applying it to a new formulation is inference.

The model is not the experiment. A formulation outside the training range, or a different manufacturing process, may invalidate its prediction. Neural networks are one family of models, and deep learning uses networks with multiple processing layers. AI is a wider field than either deep learning or conversational software.

Background: [3]

Learn from samples, then test on other samples
Figure 1. Learn from samples, then test on other samplesThe testing machine, samples and prediction screen connect measurement with modelling. Screen predictions are illustrative, not reported results. Training estimates relationships from data; it does not license claims about every possible material.Conceptual ImageGen illustration, not a photograph or measured result. Dimensions and colours are explanatory.Reused from: Machine Learning and Neural Network Fundamentals

02Predicting, generating and acting are different jobs

A classifier chooses a category. Generative AI produces text, images or other outputs from learned patterns and supplied input. A large language model, or LLM, builds on predicting continuations of sequences of small text units. Fluent wording does not mean that its factual claims have been verified.

Retrieval-augmented generation, or RAG, supplies retrieved material to help ground an answer. An agent combines a model with tools such as search or software controls to carry out work. Neither removes error automatically: the wrong document can be retrieved, and an action can be mistaken. Source checking and limited permissions remain important.

Background: [2]

A language service depends on compute, memory and connections
Figure 2. A language service depends on compute, memory and connectionsThe rack and chips illustrate the physical infrastructure behind generated text. Appearance alone establishes neither model capability nor accuracy or power consumption.Conceptual ImageGen illustration, not a photograph or measured result. Dimensions and colours are explanatory.Reused from: Transformers and Large Language Models

03Ask whether it works on genuinely unseen cases

Testing on the same questions used for practice exaggerates ability. Keep evaluation examples separate from training, and prevent answer information or future information from leaking into the input. Inspect not just aggregate accuracy but also important misses, difficult conditions and the consequences of failure.

In factory inspection, missing a defect and rejecting a good part cause different losses. A new camera or different lighting can change results without any change in the model. Deployment therefore needs monitoring and a way for people to stop the system or take over a decision.

Background: [1]

Check a proposal against sources and physical evidence
Figure 3. Check a proposal against sources and physical evidenceThe person compares a screen proposal with samples and records. An AI output needs to be checked against evidence appropriate to the task, rather than accepted simply because it is neatly presented.Conceptual ImageGen illustration, not a photograph or measured result. Dimensions and colours are explanatory.Reused from: Deploying AI in Enterprises

04Behind the model are data and physical infrastructure

Training and inference require computation, memory, communications, electricity and cooling. Accelerators perform many operations in parallel; memory and interconnects help move the required data. More infrastructure does not automatically create more business value.

Can confidential data leave the organization? How much delay is acceptable? Who checks a failure? Those questions help determine whether to use cloud services or local hardware, and a large general model or a smaller specialized one. Start with the work or decision to improve, not with a model name.

Background: [1]

05Three questions for reading an AI claim

Ask: compared with what, tested on which data, and covering how much of the real workflow? A high benchmark score, one successful demonstration and an improvement in actual work including cost and waiting time are different kinds of evidence.

A small trial against the existing workflow under comparable conditions reveals strengths and weaknesses. Decide how correctness, safety, cost and response time matter for that task. The individual articles below can then answer a concrete question rather than simply add terminology.

Background: [1]

Connect the ideas in one situation

ILLUSTRATIVE SCENARIO — NOT A PERFORMANCE CLAIM

Suppose a factory wants to find defects

First define the defects that matter and make image capture and labelling consistent. Test on products or capture dates kept outside training, and count both misses and false alarms. If uneven lighting causes failure, improving the camera setup may help more than enlarging the model.

Send uncertain cases to people after deployment. The useful role may be narrowing what needs inspection rather than eliminating every human decision. This is a workflow example, not a promised result.

Return to the starting idea

AI is a tool to match to a task and its data. A plausible output and an output that is correct for the intended use are different things.

You do not need to memorize every technology name. Understanding the intended improvement, the mechanism being changed and the conditions needed to demonstrate it gives you a foundation for the detailed articles. Choose one route that matches your question.

Sources and scope

This introductory synthesis connects the existing technical articles. Its grouping and design scenarios are explanatory, not standard classifications or test results. It does not catalogue current records, product availability or approvals, or forecast costs. Foundational references are below; linked articles provide more detailed technical sourcing.

  1. NIST. AI Risk Management Framework
  2. Vaswani et al.. Attention Is All You Need
  3. Goodfellow, Bengio & Courville. Deep Learning
← Back to this field’s guide library
🌐 Japanese