Using the force: STEM knowledge and experience construct shared neural representations of engineering concepts

The study investigates how classroom and lab-based STEM learning shapes neural representations of abstract mechanical concepts when viewing naturalistic stimuli. Grounded in prior findings that experts categorize problems by underlying principles while novices rely on surface features, the authors ask how abstract physics/engineering knowledge is implicitly represented in the brain. Prior neuroimaging work has focused on explicit retrieval with non-naturalistic stimuli, implicating dorsal stream and motor regions in task-specific physics knowledge. In contrast, many studies link visual categorization (including implicit categorization) to ventral occipitotemporal cortex (vOT). Using a cross-sectional design, the authors compare mechanical engineering students to novice peers while both perform an fMRI free body diagram (FBD) task, evaluating forces on real-world structures. They apply multivariate pattern analysis, intersubject correlation, and representational similarity analyses to test for group-specific convergent neural representations and whether these reflect mechanical category knowledge beyond visual similarity. The goal is to reveal how prior STEM learning constructs neural patterns that encode abstract mechanical categories during naturalistic perception.

The paper situates its work within: (1) cognitive psychology showing expert–novice differences in categorization of physics problems by deep principles versus surface features; (2) neuroimaging studies of category representation in vOT demonstrating implicit conceptual categorization of real-world stimulus classes; and (3) limited prior neuroimaging on physics/engineering knowledge that mainly examined explicit retrieval with artificial stimuli and implicated dorsal stream and motor regions (e.g., M1) for concepts like torque and causal motion. The present work extends these by probing implicit activation of abstract mechanical categories from real-world images during an incidental force-evaluation task, and by dissociating conceptual content from bottom-up visual similarity using a control model (HMAX).

Design: Cross-sectional fMRI study comparing mechanical engineering undergraduates (with lecture and hands-on lab experience; advanced physics prerequisite) to novice undergraduates without advanced engineering/physics courses. Participants: N=31 (19 female; mean age=20.65, SD=1.70). Groups: novices (N=15) and engineering students (N=16). Stimuli: 24 photographs of engineered structures. Each image had a component highlighted (red outline) for force analysis; two labeled versions (correct/incorrect) indicated forces/moments for decision-making. Task (FBD): In the scanner, each trial: 2 s image; 4 s with highlighted component; jittered fixation during which participants imagined forces/moments for static equilibrium (first 6 s of trial comprised the analyzed period); then 4 s with labels (correct/incorrect) to judge via button press; additional jittered fixation; trials 15.5 s with 15.5 s inter-trial baseline; 4 runs, each including all 24 images; 50% correctly labeled per run, counterbalanced across runs. Procedure: Behavioral session (similarity ratings; SCI and FCI tests—results not discussed). fMRI session: concept primer on Newtonian force, static equilibrium, and FBDs; familiarization and practice with feedback (experimental runs without feedback); four runs total. Imaging: 3T Philips Achieva Intera, 32-channel head coil; gradient-echo EPI; 80x80 matrix; FOV 240 mm; 42 transverse slices; flip angle 90°; TE=35 ms; TR=2500 ms; 3 mm isotropic voxels; no gap; 298 volumes per run. High-res T1: TE=3.72 ms; TR=8.176 ms; 0.938×0.938×1.0 mm. Preprocessing (FSL FEAT): skull stripping (BET), motion correction, slice timing correction, prewhitening, high-pass filter (100 s), registration to anatomy. GLM: item-level regressors modeling first 6 s (perception/force imagination); separate regressor for response period; beta estimates from initial 6 s used for MVPA. Surface-based analysis: Freesurfer recon-all; SUMA; surfaces fitted to icosahedral mesh (32 divisions; 20,484 nodes); sulcal alignment to MNI template for correspondence. Multivariate analysis: Step 1: For each subject and node, compute item-level dissimilarity matrices (DMs) via pairwise correlations of item-wise beta patterns (PyMVPA). Step 2: Intersubject DM correlations within group at each node; Fisher z-transform; average z per node; threshold using negative extent as noise estimate (z>0.02), validated with permuted nulls; spatial cluster correction on cortical surface (≥5 contiguous nodes). Run 1 emphasized due to maximal prior-knowledge variance; Run 4 also analyzed for comparison. Step 3: Informational network analysis: Within distinct/overlapping regions from Step 2, perform Ward hierarchical clustering of average node-level DMs to define informational networks; number of clusters chosen via split-half cross-validation (2–100 clusters; 1000 repetitions). For each participant, compute average DM per informational network. RSA: correlate network DMs with a mechanical category model DM (expert-defined categories: cantilevers, trusses, vertical loads); one-sample t-tests across participants per network to yield group-level t-maps of mechanical information. Step 4: Visual similarity control: compute HMAX C1-layer model DM; repeat RSA over same networks; compare spatial gradients and peaks with mechanical model. Analyses prioritize Run 1; Run 4 used to assess task-specific learning effects.

Behavior: Engineering students outperformed novices on the FBD task overall (Mean_eng=76%, Mean_nov=66%, t(29)=2.44, p=0.02). Linear mixed-effects: significant effects of run (β=0.03, SE=0.01, p=0.01) and group (engineering > novice; β=0.24, SE=0.06, p=0.0005), with a run×group interaction (β=0.05, SE=0.01, p=0.0005): novices improved more over runs (Δ_eng=7.55%, Δ_nov=23.61%, t(29)=3.47, p=0.002). Largest group difference at run 1 (Mean_eng=74%, Mean_nov=53%, t(29)=4.18, p=0.0002); by run 4, no significant difference (t(29)=1.34, p=0.19). Neural—Run 1 intersubject convergence: Of 20,484 surface nodes, 1,863 nodes showed convergence only in engineering students, 842 only in novices, and 2,426 in both. Engineering-specific convergence localized to M1, ventral PFC, and inferior parietal regions. Novice-specific convergence localized to anterior frontal and dorsal parietal regions. Overlap occurred in occipital, dorsal occipital, and vOT, with distinctions in most anterior vOT. Mechanical category information (RSA over informational networks): Engineering students showed a significant peak correlation with the mechanical category model in bilateral anterior vOT (t(15)=2.27, p=0.038) and a nonsignificant secondary peak in M1 (t(15)=1.93, p=0.073). Novices showed a nonsignificant mechanical peak in right anterior vOT (t(14)=2.06, p=0.058). Peaks for each group arose from group-specific convergent regions (engineering: bilateral anterior vOT and left M1 within engineer-specific convergence; novice: right anterior vOT within novice-specific convergence). Gradients: Mechanical information increased posterior-to-anterior along vOT in both groups (significant in engineers) and along the dorsal stream toward M1 in engineers. Visual similarity control (HMAX): Both groups showed strong visual similarity information peaking in V1 (engineers: t(15)=8.92, p<0.001; novices: t(14)=6.48, p<0.001) within overlapping convergence regions. Visual information increased anterior-to-posterior along vOT—opposite direction of the mechanical category gradient—indicating dissociable representations and ruling out low-level visual similarity as the basis for mechanical category effects. Run 4: Fewer convergent regions (15 unique networks engineers; 2 unique novices; 43 overlapping), mostly confined to vOT with one right posterior parietal region per group; no convergent dorsal premotor or M1 representations; RSA showed weaker mechanical category information but strong visual feature information persisted. This pattern suggests run-to-run changes reflect task-specific learning rather than changes in conceptual knowledge.

Findings indicate that prior classroom and lab-based engineering education shapes neural representations engaged during naturalistic force reasoning. Engineering students exhibited group-convergent multivariate patterns that encoded abstract mechanical categories, despite categories never being mentioned. Novices, with poorer initial task performance, lacked significant mechanical category representations at run 1. Mechanical category information localized predominantly to bilateral anterior vOT and showed a dorsal gradient toward M1 in engineers, aligning with prior reports of ventral stream conceptual coding and dorsal/motor involvement in physics concepts. The dissociation from visual similarity (opposite gradients; V1 peaks shared across groups) supports that conceptual mechanical information cannot be reduced to low-level visual features. By run 4, convergent representations were largely confined to vOT and reflected visual features more than mechanical categories, consistent with task-specific familiarity rather than conceptual change. Overall, the work demonstrates that abstract STEM concepts learned in formal education are implicitly activated by real-world stimuli and are detectable in distributed neural patterns, with group-specific localization reflecting learned categories.

The study shows that engineering education constructs shared neural representations of abstract mechanical categories in bilateral anterior vOT and, to a lesser extent, dorsal/motor regions, when students evaluate forces on real-world structures. These representations are distinct from visual similarity encoding and emerge without explicit category instruction, indicating implicit activation of learned STEM concepts. Future research could longitudinally track how such conceptual representations develop with instruction, examine causality and the role of hands-on experience (e.g., M1 involvement), expand to other STEM domains and categories, and test instructional interventions that enhance conceptual encoding in ventral and dorsal streams.

The study is cross-sectional, limiting causal inference about learning-induced representational change. Primary neural inferences are drawn from the first fMRI run to capture prior-knowledge differences; later runs show task-specific learning effects, not conceptual change. Mechanical categories were defined by expert judgment, which could bias the target model. The visual control relied on the HMAX C1 model, which may not capture all perceptual similarities. The sample size (N=31) is modest, and correspondences at run 4 were sparse outside vOT, limiting generalizability across tasks and time. Corresponding author and some affiliations were not fully detailed in the excerpt.