Steel: Alloys, Microstructure & Applications

• Steel is a family of ferrous alloys defined by controlled composition, processing routes, and microstructural features, offering tailored strength, ductility, and corrosion resistance.
• Recent studies employ machine learning and precise heat treatment methods to optimize alloy compositions and predict key mechanical properties.
• Advanced inspection, joining, and surface engineering techniques, including robotics and non-destructive evaluation, are enhancing steel manufacturing and reliability.

Searching arXiv for the supplied steel-related papers and a recent overview to ground the article in current arXiv records. Steel is a family of ferrous alloys whose engineering behavior is controlled by composition, processing, microstructure, and service environment. In the broad Fe–C framework, steels extend up to roughly $2\ \mathrm{wt.\%}$ C, in contrast to cast irons at typically $3$–$4.5\ \mathrm{wt.\%}$ C, and their major microstructural constituents include ferrite, austenite, pearlite, bainite, martensite, and tempered martensite (Sgobba, 3 Sep 2025). Stainless steel is conventionally defined as a ferrous alloy containing at least $12\%$ Cr, and the technologically dominant stainless families are ferritic, martensitic, austenitic, duplex, and precipitation-hardening grades (Sgobba, 3 Sep 2025). Across structural, cryogenic, vacuum, magnetic, electrochemical, and manufacturing applications, steel is therefore not a single material but a process-dependent class of alloys whose usable properties emerge from tightly specified chemistry, steelmaking route, cleanliness, hot working, heat treatment, joining, and surface condition (Sgobba, 3 Sep 2025).

1. Classification, constitution, and alloy design

The central metallurgical distinction in the supplied literature is between ordinary carbon and low-alloy steels, high-Mn advanced high-strength steels, high-Si electrical steels, stainless steels, and specialized cryogenic or vacuum-compatible grades. Carbon steels and low-alloy steels are treated explicitly in a machine-learning study based on 360 heat-treated records from the NIMS structural materials database, where the relevant variables included three heat-treatment temperatures, nine alloying elements, one reduction ratio, and three non-metallic inclusion measures (Xiong et al., 2020). In that dataset the principal room-temperature properties were fatigue strength, tensile strength, fracture strength, and hardness, and the dominant features selected for compact modeling were tempering temperature, carbon, chromium, and molybdenum (Xiong et al., 2020).

Stainless steels are differentiated in the data primarily by phase stability and service function. Ferritic stainless steels, typically $14.5$–$27\%$ Cr, remain ferritic up to melting, are ferromagnetic, and exhibit a ductile-to-brittle transition, which limits cryogenic applicability (Sgobba, 3 Sep 2025). Martensitic stainless steels rely on high-temperature austenite formation and quench-induced martensite, offering high strength at the cost of ferromagnetism and lower cryogenic suitability (Sgobba, 3 Sep 2025). Austenitic stainless steels, stabilized by Ni, Mn, and N, are the preferred class for accelerator, fusion, cryogenic, and vacuum systems because they are paramagnetic, weldable, corrosion resistant, and do not show the ferritic DBTT behavior (Sgobba, 3 Sep 2025). Within this class, the paper emphasizes 304L, 316L, 316LN, 316Ti, P506, and FXM-19/Nitronic 50, each selected by balancing austenite stability, low magnetism, toughness, vacuum compatibility, and fabrication constraints (Sgobba, 3 Sep 2025).

High-Mn advanced high-strength steels in the supplied papers illustrate how composition is used to tune deformation mechanisms rather than only static strength. A Fe-0.07C-2.85Si-15.3Mn-2.4Al-0.017N steel exhibited two sequential TRIP mechanisms, with initial $\gamma \rightarrow \epsilon$ transformation followed by $\epsilon \rightarrow \alpha$, producing a maximum strain-hardening exponent of $1.4$, ultimate tensile strength of $1165\ \text{MPa}$, and elongation to failure of $3$0 (McGrath et al., 2012). A different medium-Mn steel, Fe-12Mn-4.8Al-2Si-0.32C-0.3V, was engineered to develop a necklace-core microstructure and reached $3$1 yield strength, $3$2 ultimate tensile strength, and $3$3 elongation (Kwok et al., 2019). In both cases, austenite stability, stacking-fault energetics, and partitioning of Mn, Al, Si, and C determined whether deformation proceeded by TRIP, TWIP, or both (McGrath et al., 2012, Kwok et al., 2019).

High-silicon electrical steel appears as another distinct steel class. Fe–6.5 wt% Si is identified as a premium soft magnetic steel because of high saturation magnetization, high resistivity, low iron loss, and near-zero magnetostriction, but its industrial use is limited by brittleness and processing difficulty (Xu et al., 2024). The paper on double glow plasma surface metallurgy specifically targets conversion of a Fe–3.6 wt% Si sheet into either gradient or nearly homogeneous Fe–6.5 wt% Si sheet by plasma-assisted siliconizing (Xu et al., 2024).

2. Microstructure and deformation mechanisms

The supplied studies consistently show that steel behavior is microstructure-mediated. In carbon and low-alloy steels, lower tempering temperature and higher contents of carbon, chromium, and molybdenum tend to increase fatigue strength, tensile strength, fracture strength, and hardness, with chromium showing curvature and molybdenum contributing positively within the studied range (Xiong et al., 2020). The symbolic-regression relations reported for the 360-steel NIMS dataset encode these trends explicitly. For example, the fitted fatigue-strength relation is

$3$4

with analogous expressions for tensile strength, fracture strength, and hardness (Xiong et al., 2020). These formulas are less accurate than the best black-box models but expose the directional role of tempering temperature and strengthening alloy additions (Xiong et al., 2020).

In the high-Mn TRIP steel Fe-0.07C-2.85Si-15.3Mn-2.4Al-0.017N, the starting ferrite-free microstructure contained about $3$5 retained austenite, $3$6 $3$7-martensite, and $3$8 $3$9-martensite (McGrath et al., 2012). During the first $4.5\ \mathrm{wt.\%}$0 strain, retained austenite transformed primarily to $4.5\ \mathrm{wt.\%}$1-martensite, and only after saturation of this stage did $4.5\ \mathrm{wt.\%}$2-martensite transform strongly to $4.5\ \mathrm{wt.\%}$3-martensite, which coincided with the rapid rise in strain hardening (McGrath et al., 2012). The mechanistic explanation explicitly invokes both intrinsic stacking fault energy and unstable stacking fault energy: Al lowers $4.5\ \mathrm{wt.\%}$4, promoting easy nucleation of $4.5\ \mathrm{wt.\%}$5-martensite, while raising $4.5\ \mathrm{wt.\%}$6, rendering $4.5\ \mathrm{wt.\%}$7-martensite unstable enough to transform onward to $4.5\ \mathrm{wt.\%}$8-martensite under continued deformation (McGrath et al., 2012). Segregation broadened the transformation window further by leaving Mn-rich regions untransformed until later strain, thereby extending work hardening near necking (McGrath et al., 2012).

The medium-Mn steel Fe-12Mn-4.8Al-2Si-0.32C-0.3V uses a different microstructural strategy: coarse elongated core austenite grains, larger than $4.5\ \mathrm{wt.\%}$9, are surrounded by necklace layers of fine austenite and ferrite grains much smaller than $12\%$0 (Kwok et al., 2019). The coarse core austenite, being less stable, transforms first by TRIP and supplies the initial hardening stage; the finer, more stable necklace austenite then contributes a later stage through TWIP plus TRIP, confirmed by TEM and in-situ neutron diffraction (Kwok et al., 2019). This hierarchical sequence sustained a near-constant hardening rate of $12\%$1 in the second stage and delayed necking to unusually high strain (Kwok et al., 2019). A plausible implication is that controlled heterogeneity in austenite morphology and chemistry can be as important as average phase fraction in advanced steel design.

In stainless steels for cryogenic and vacuum service, the desired microstructure is substantially or fully austenitic. The overview paper treats $12\%$2-ferrite, martensite formed by deformation or cooling, and grain-boundary precipitates as undesirable because they raise magnetic permeability, lower cryogenic ductility, or reduce corrosion and vacuum reliability (Sgobba, 3 Sep 2025). The same paper emphasizes fine and homogeneous grain size, low segregation, and very low inclusion content as necessary microstructural conditions for leak-tight and non-magnetic service (Sgobba, 3 Sep 2025).

3. Processing, joining, and surface engineering

The article set makes clear that steel properties are inseparable from process route. The machine-learning study on carbon and low-alloy steels used 16 descriptors that included not only composition but also normalizing, quenching, and tempering temperatures, reduction ratio, and inclusion descriptors (Xiong et al., 2020). In those data, the omission of heating and cooling rates and detailed holding times was already identified as a limitation on generalization (Xiong et al., 2020). For the medium-Mn steel, the thermomechanical route consisted of vacuum arc melting, homogenization at $12\%$3 for 24 h, hot rolling at $12\%$4, warm rolling at $12\%$5, and a 30 min intercritical anneal at $12\%$6, yielding a continuous industrially translatable schedule without quench interruption or cold rolling (Kwok et al., 2019). In the two-stage TRIP steel, processing involved homogenization at $12\%$7, hot rolling from $12\%$8 with reheats, reduction of $12\%$9 to $14.5$0, and water quenching after a final $14.5$1, 10 min treatment (McGrath et al., 2012).

Processing route also governs whether highly functional steels can be made at all. The Fe–6.5 wt% Si electrical-steel paper argues that direct rolling of high-silicon steel is obstructed by severe brittleness, and instead demonstrates double glow plasma surface metallurgy using a monocrystalline silicon source cathode, argon pressure $14.5$2–$14.5$3, source voltage $14.5$4–$14.5$5, work-piece voltage $14.5$6–$14.5$7, alloying temperature $14.5$8–$14.5$9, and holding time $27\%$0–$27\%$1 (Xu et al., 2024). With a 6 h treatment, a $27\%$2 Fe–3.6 wt% Si sheet reached near-homogeneous through-thickness silicon levels of $27\%$3, $27\%$4, and $27\%$5 from one surface to the center and achieved iron loss at 50 Hz of $27\%$6, down from $27\%$7–$27\%$8 in the untreated sheet (Xu et al., 2024).

Joining dissimilar steels is addressed by directed energy deposition of a graded SS316L-to-C300 maraging steel transition. The build used 100 layers of pure C300, a 13-layer graded region with 7.143 vol.% composition increments, and 100 layers of pure SS316L, deposited at $27\%$9 scan speed, $\gamma \rightarrow \epsilon$0 layer thickness, and $\gamma \rightarrow \epsilon$1 hatch width (Ben-Artzy et al., 2021). Remelting produced an almost continuous chemistry transition over about $\gamma \rightarrow \epsilon$2–$\gamma \rightarrow \epsilon$3, with EBSD showing a progression from BCC martensitic structure on the maraging side through FCC/BCC dual-phase material to fully FCC austenite on the stainless side (Ben-Artzy et al., 2021). No intermetallic phases were detected in the interface region, and the as-built interface strength was essentially similar to the stainless side, although elongation to fracture was reduced (Ben-Artzy et al., 2021).

Surface engineering appears in two very different contexts. One is galvanizing of second-generation AHSS. In nano-TWIP steel of nominal bulk composition about $\gamma \rightarrow \epsilon$4 Fe, $\gamma \rightarrow \epsilon$5 Mn, and $\gamma \rightarrow \epsilon$6 C, low-dew-point annealing led to manganese segregation and formation of a continuous $\gamma \rightarrow \epsilon$7–$\gamma \rightarrow \epsilon$8 external MnO layer that blocked zinc wettability and the normal $\gamma \rightarrow \epsilon$9 inhibition layer (Arndt et al., 2013). The result was discontinuous galvanizing, with coating success only where Fe remained locally exposed or where isolated Fe-rich platelets avoided full MnO coverage (Arndt et al., 2013). The other is the electrochemical oxidation of Ni42 steel in 4.8 M LiOH, which converted an initially unstable Fe/Ni/Mn alloy surface into a Ni-enriched Fe/Ni oxyhydroxide-containing outer zone suitable for acidic oxygen evolution (Schäfer et al., 2018). Under optimized conditions, the Ni42Li205 electrode reached an overpotential of $\epsilon \rightarrow \alpha$0 at $\epsilon \rightarrow \alpha$1 in 0.5 M $\epsilon \rightarrow \alpha$2 and showed mass loss of only $\epsilon \rightarrow \alpha$3 after $\epsilon \rightarrow \alpha$4 at pH 1, compared with $\epsilon \rightarrow \alpha$5 for untreated Ni42 (Schäfer et al., 2018).

4. Functional properties and domain-specific applications

The application range represented in the supplied papers is unusually broad. Structural and fatigue-critical steels are treated through the NIMS-based property-prediction study, where fatigue strength is defined as rotating-bending fatigue strength at $\epsilon \rightarrow \alpha$6 cycles and paired with tensile strength, fracture strength, and hardness as central room-temperature metrics (Xiong et al., 2020). The authors’ practical design example chose the lowest tempering temperature in the dataset, $\epsilon \rightarrow \alpha$7, and the maximum values of $\epsilon \rightarrow \alpha$8, $\epsilon \rightarrow \alpha$9, and $1.4$0, yielding predicted properties of fatigue strength $1.4$1 at $1.4$2 cycles, tensile strength $1.4$3, fracture strength $1.4$4, and hardness $1.4$5 (Xiong et al., 2020). This suggests that interpretable steel design can be embedded into search-space reduction workflows even when the data are modest.

Electrical and magnetic applications motivate both high-silicon steel and specialized stainless steels. Fe–6.5 wt% Si is described as attractive for motor, generator, and transformer cores because high resistivity lowers eddy-current loss, iron loss is low, and magnetostriction is nearly zero (Xu et al., 2024). In accelerator and fusion systems, the relevant magnetic requirement is often the opposite: austenitic stainless steels are chosen to minimize ferromagnetism. The overview paper notes that grades like 304L can develop increased magnetic susceptibility through martensitic transformation or residual $1.4$6-ferrite, while P506 was specifically developed to maintain relative magnetic permeability below 1.005 at cryogenic temperature in both base metal and welds (Sgobba, 3 Sep 2025).

Cryogenic and vacuum service constitute another major application domain. The overview paper places austenitic stainless steels at the center of large particle accelerators, detectors, fusion reactors, superconducting magnets, and associated vacuum and cryogenic systems because they combine leak-tightness, low magnetism, corrosion resistance, and superior toughness down to helium temperatures (Sgobba, 3 Sep 2025). Example applications include 316LN seamless tube for LHC cold bores, 316LN plates for the shrinking cylinder of dipole magnets, HIPed 316LN powders for magnet end covers, and specially remelted 316L thin sheet for bellows convolutions operating at $1.4$7 (Sgobba, 3 Sep 2025). A separate study on vacuum materials for future gravitational-wave detectors broadens the picture by evaluating mild steels as low-cost UHV beam-pipe candidates. After bakeouts at $1.4$8 for 48 h, the tested mild steels showed hydrogen outgassing below $1.4$9, often in the $1165\ \text{MPa}$0 range, which the authors compare favorably to ordinary 304L after similar low-temperature baking (Scarcia et al., 2024). The caveat is that after such bakeouts the dominant residual-gas limitation is water vapor, strongly dependent on oxide/hydroxide surface condition (Scarcia et al., 2024).

Steel also appears as a material platform for electrochemistry and energy systems. The Ni42 acidic-OER study shows that a steel substrate can function not merely as a support but as a steel-derived catalytic electrode after suitable electrochemical surface modification (Schäfer et al., 2018). By contrast, the CO$1165\ \text{MPa}$1-steelmaking proposal uses steel at the process level rather than as end-product function, suggesting decarburization of hot metal by the Boudouard reaction,

$1165\ \text{MPa}$2

with external heating and recycle of regenerated CO$1165\ \text{MPa}$3 (Ramayya et al., 2015). The evidence provided is only a lab-scale decarburization of high-carbon ferro-chrome powder from $1165\ \text{MPa}$4 to $1165\ \text{MPa}$5 carbon at $1165\ \text{MPa}$6 over 24 h under pure CO$1165\ \text{MPa}$7, so the proposal remains conceptual for bulk molten steel (Ramayya et al., 2015).

5. Inspection, defect detection, and data-centric steel engineering

Steel research in the supplied corpus includes not only metallurgy but also inspection robotics, nondestructive evaluation, computer vision, and datasets. A climbing-robot study developed a four-wheel magnetically adhered robot for steel structures and bridges, with two video cameras, a ToF camera, Hall-effect sensors, and IR sensors, capable of adhering to inclined and upside-down steel surfaces while capturing images for stitching and 3D reconstruction (La, 2017). The robot used permanent magnetic adhesion and was tested under multiple conditions, including coated, rusty, curved, and vertical steel surfaces (La, 2017). A related tank-like bridge robot used magnetic roller-chains and a reciprocating transformation mechanism to adapt to flat and cylindrical members, reportedly deploying on more than 20 steel bridges and handling bolts, nuts, and geometry transitions (Nguyen et al., 2018). Both works treat steel inspection as a coupled locomotion-sensing problem, not only an image-analysis problem.

The sensing modality itself is varied. One paper demonstrates nondestructive inspection of 316 stainless steel using a fiber-coupled nitrogen-vacancy-center diamond magnetometer in a deliberately inhomogeneous bias field (Zhou et al., 2020). Structural damage perturbs that field, and the system reconstructs slot depth or width through shifts in NVC Zeeman splitting (Zhou et al., 2020). The achieved performance was about $1165\ \text{MPa}$8 lateral spatial resolution, $1165\ \text{MPa}$9 perpendicular resolution, lift-off up to $3$00, and successful imaging through $3$01 brass or $3$02 fiberglass coverings (Zhou et al., 2020). The paper positions this as distinct from classical magnetic flux leakage because it does not require driving the steel close to magnetic saturation (Zhou et al., 2020).

Computer-vision inspection is represented at several scales. A one-class anomaly-detection system for steel surfaces tiles large images into $3$03 unit images, classifies region of interest versus background, extracts discriminator features from a PatchGAN trained on normal steel imagery, and scores anomalies with a one-class SVM (Yasuno et al., 2021). It was demonstrated on 13,774 unit images from steel-sheet inspection and 19,766 unit images from painted-steel corrosion inspection (Yasuno et al., 2021). A more conventional supervised detector based on YOLOv9s, SCConv, C3Ghost, and CARAFE was evaluated on the NEU steel surface defect dataset of 1,800 images across six classes—Crazing, Inclusion, Patches, Pitted Surface, Rolled-in Scale, and Scratches—and improved [email protected] from $3$04 for baseline YOLOv9s to $3$05, while reducing parameters from $3$06 M to $3$07 M and increasing FPS from 94.3 to 109.9 (Chen et al., 21 Jul 2025). The most notable per-class gains were on Inclusion, Crazing, and Rolled-in Scale, although recall dropped relative to baseline (Chen et al., 21 Jul 2025).

At the dataset level, SteelDS provides 24,297 labeled frames of shredded E40-grade steel and copper scrap on a conveyor belt, with 396 steel objects and 101 copper objects, instance-segmentation masks, and controlled spacing/occlusion regimes (Neubauer et al., 26 May 2026). The benchmark is explicitly designed for post-magnetic sorting, where copper contamination remains in a predominantly steel stream (Neubauer et al., 26 May 2026). In combined fitness, Mask R-CNN outperformed several YOLO nano segmentation models, and medium-sized objects were substantially harder than large ones for both detection and segmentation (Neubauer et al., 26 May 2026). A plausible implication is that for steel recycling, exact segmentation of irregular metallic fragments is limited more by clutter, scale, and topology than by class count.

Physics-informed and interpretable data-centric approaches also appear in bulk steel design. The physics-informed CCT framework trained on 4,100 diagrams and 35,820 cooling-rate-specific data points predicts ferrite, pearlite, bainite, and martensite formation under continuous cooling, generating complete diagrams with 100 cooling curves in under 5 s and achieving phase-classification F1 scores above 88% for all phases (Hedström et al., 21 Nov 2025). Reported regression errors were below $3$08 for all phases except bainite, which was around $3$09 in the abstract summary (Hedström et al., 21 Nov 2025). This places steel among the clearest examples where metallurgical prior knowledge—here encoded through start-temperature coupling, phase-fraction constraints, and synthetic low-cooling-rate anchors—materially improves ML utility (Hedström et al., 21 Nov 2025).

6. Failure mechanisms, constraints, and open engineering issues

Despite the breadth of successful applications, the supplied literature repeatedly shows that steel performance is conditional and failure-prone when chemistry, process, or surface state are not controlled. In galvanizing of high-Mn AHSS, the primary failure mechanism is selective oxidation: continuous external MnO blocks wetting by Zn and prevents formation of the $3$10 inhibition layer (Arndt et al., 2013). In acidic electrolysis, untreated Ni42 steel is electrochemically active but corrodes rapidly, losing $3$11 after $3$12 at $3$13 in 0.05 M $3$14 (Schäfer et al., 2018). In mild-steel vacuum tubing, hydrogen is not the dominant obstacle after low-temperature bakeout; instead, water retained or regenerated by oxide/hydroxide surface layers controls the ultimate pressure, and poor industrial oxide morphology can generate virtual-leak behavior, as in the S355J2H tube with a roughly $3$15 cracked oxide layer (Scarcia et al., 2024).

Advanced austenitic stainless steels also have failure modes tied to phase stability and thermal exposure. The overview paper warns that residual $3$16-ferrite, deformation-induced martensite, grain-boundary carbide precipitation, and sensitization can simultaneously compromise magnetic, vacuum, corrosion, and cryogenic requirements (Sgobba, 3 Sep 2025). In one ITER 316LN jacket example, sensitization and associated microstructural changes reduced elongation at 7 K to $3$17, and only stricter chemistry and process control restored elongation to at least $3$18 (Sgobba, 3 Sep 2025). Inclusion cleanliness is another recurring concern: for vacuum-tight components, aligned inclusions elongated by working can generate through-thickness leak paths, so remelting, multidirectional forging, and fiber-orientation control are treated as design variables, not merely procurement details (Sgobba, 3 Sep 2025).

Data-driven steel methods have their own limitations. The NIMS ML study is based on only 360 records after filtering, excludes stainless steels, and lacks heating/cooling rates and most holding-time details (Xiong et al., 2020). The symbolic formulas are interpretable but less accurate than the best RF and ANN models (Xiong et al., 2020). The CCT framework, while large-scale, still omits proeutectoid cementite and some high-alloy complexities, and bainite remains the hardest regime because of its strong dependence on prior transformation history and ppm-level alloying effects such as B (Hedström et al., 21 Nov 2025). Computer-vision systems remain sensitive to very small defects, strong illumination variation, or complex surface textures (Chen et al., 21 Jul 2025, Yasuno et al., 2021). Inspection robotics, although effective, still faces lighting sensitivity, geometric constraints, and incomplete autonomy on steel bridges (La, 2017, Nguyen et al., 2018).

A broad conclusion across the papers is that steel selection cannot be reduced to nominal grade names. The supplied overview on steels and stainless steels makes this explicit for large scientific infrastructure: steel products and grades must be specified together with microstructure, cleanliness, inspection route, and final application constraints across vacuum, magnetic, and cryogenic domains (Sgobba, 3 Sep 2025). The supplied applied papers reinforce the same point in different language. Whether the target is fatigue strength, cryogenic permeability, galvanizability, beam-pipe outgassing, acidic electrocatalysis, additive joining, or defect detectability, the decisive variable is usually not “steel” in the abstract but a tightly defined steel state produced by a particular process history (Xiong et al., 2020, Arndt et al., 2013, Schäfer et al., 2018, Ben-Artzy et al., 2021, Scarcia et al., 2024).