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

The study explores how prior learning experiences influence the neural representation of abstract conceptual categories when presented with naturalistic stimuli. Existing research shows that experts and novices categorize problems differently based on their knowledge; experts use abstract concepts while novices rely on surface features. While neuroimaging studies have focused on visually perceivable categories, the neural representation of abstract STEM concepts remains less explored. Few neuroimaging studies have investigated physics and engineering concept knowledge, with existing studies primarily focusing on explicit retrieval of task-specific knowledge from non-naturalistic stimuli. This study aimed to resolve how abstract knowledge is implicitly activated by real-world stimuli. Utilizing fMRI, the researchers compared undergraduate mechanical engineering students (with classroom and lab experience) to a control group of novice students matched for educational attainment. Both groups performed an fMRI concept knowledge task evaluating Newtonian forces acting on real-world structures. Multivariate pattern analysis (MVPA) and informational network analysis (a variant of representational similarity analysis, RSA) were used to identify patterns of brain activity associated with real-world stimuli that implicitly reflect learned abstract mechanical category information.

Seminal work in cognitive psychology has shown that experts and novices differ in how they categorize problems. Experts use abstract conceptual knowledge while novices focus on surface-level features. While many fMRI studies have linked visual categorization to ventral visual stream activity, the neural representation of abstract STEM concepts is less explored. Previous neuroimaging studies on physics and engineering concept knowledge have implicated dorsal stream regions, including motor cortex, in the explicit retrieval of task-specific knowledge. However, these studies used non-naturalistic stimuli, leaving unresolved how abstract knowledge is implicitly activated by real-world stimuli. This study builds on previous research using similar analytical approaches in other concept domains, where patterns of neural activity in ventral occipito-temporal cortex reliably classify images into abstract categories. The researchers adapted this approach to the domain of mechanical engineering.

Thirty-one Dartmouth College students participated (15 engineering students, 16 novices). Stimuli consisted of 24 photographs of real-world engineered structures. Participants performed a free body diagram (FBD) task during fMRI scanning, evaluating Newtonian forces acting on these structures. The task involved three stages: (1) initial viewing of the structure, (2) a period of mental analysis to imagine the forces, and (3) viewing the structure with forces labeled correctly or incorrectly, requiring a judgment from the participant. fMRI data from the mental analysis period (first 6 seconds) were used for analysis. Before the fMRI session, participants completed a similarity rating task and standardized tests of engineering and physics knowledge (results not reported in this paper). Functional MRI data were preprocessed using FSL FEAT, including skull stripping, motion correction, and slice timing correction. Data were registered to individual anatomical volumes and then to a standard cortical surface mesh, resulting in 20,484 nodes. Multivariate pattern analysis (MVPA) was performed to identify regions showing convergent neural activity within each group. Informational network analysis (a form of RSA) was then used to determine if neural activity reflected mechanical category information (cantilevers, trusses, vertical loads). A visual similarity model (HMAX) served as a control. Data from the first and fourth fMRI runs were analyzed. Inter-subject correlation analysis was conducted to identify convergent neural representations within each group. A noise threshold (z > 0.02) was applied, with spatial clustering used to identify significant clusters. Informational network analysis used Ward hierarchical clustering to identify networks of brain regions with similar representations. These networks were then correlated with a mechanical category model (created by an expert) and a visual similarity model (HMAX). One-sample t-tests compared correlation distributions against zero to assess consistency.

Engineering students significantly outperformed novices on the FBD task, particularly during the first fMRI run, reflecting pre-existing knowledge. MVPA revealed distinct and overlapping neural activity patterns between the groups. Engineering students showed unique representational convergence in motor regions (M1), ventral PFC, and inferior parietal regions. Novices showed convergence in anterior frontal and superior parietal regions. Overlapping convergence occurred in occipital regions, including vOT. Informational network analysis showed that engineering students’ neural patterns strongly reflected mechanical category information, particularly in bilateral anterior vOT (t(15) = 2.27, p = 0.038) and, less significantly, in M1 (t(15) = 1.93, p = 0.073). Novices showed no significant mechanical category representation. The visual similarity analysis revealed that both groups showed peak visual similarity information in V1, increasing along an anterior-to-posterior gradient in the ventral stream (opposite to the mechanical category information gradient). Analysis of the fourth fMRI run showed less convergent neural activity overall, primarily in vOT. Mechanical category information was less strongly represented. These results suggest that neural representations of Newtonian force concept knowledge are best identified at the first run, reflecting pre-existing knowledge. At the final run, the representations mainly reflected task-specific learning.

The findings demonstrate that classroom-based STEM knowledge and experience directly influence neural representations accessed during naturalistic problem-solving. Engineering students implicitly accessed mechanical category knowledge while performing the FBD task, even without explicit mention of categories. The significant difference in neural activity between groups in the first fMRI run, and the decline in the difference in the fourth fMRI run, underscores the role of prior knowledge. The discrepancy between the gradient of visual and mechanical information strengthens the argument that these are distinct representations. The localization of mechanical category knowledge in bilateral anterior vOT aligns with previous research on object concepts. The involvement of M1 might be related to hands-on lab experience. This study shows that abstract conceptual knowledge acquired in school is activated in response to real-world stimuli and can be identified in multivariate neural response patterns.

This study demonstrates that prior STEM learning influences neural representations of objects during problem solving. Engineering students showed distinct neural patterns reflecting mechanical category knowledge not present in novices. The use of MVPA and RSA successfully uncovered implicit conceptual knowledge in a naturalistic context, supporting the idea that learning modifies how we perceive the world. Future research should explore the specific role of hands-on experience in shaping these neural representations.

The cross-sectional design limits causal inferences. The sample size, while adequate, could be increased for greater statistical power. The study focused on a specific domain (mechanical engineering), limiting generalizability. Task-specific learning potentially confounds the results, particularly in later fMRI runs.