Imagine training a self-driving car by showing it only footage of a professional driver perfectly centered in their lane. The car never sees a wheel slip onto the shoulder, never experiences a skid, and never witnesses a crash. When placed on a real mountain road, it will inevitably fail because it has no concept of where the asphalt ends and the cliff begins.
Machine learning models applied to the physical sciences face an identical crisis. When we train models to discover new drugs, predict chemical reactions, or design novel materials, we feed them datasets derived from published scientific literature. However, this literature suffers from a profound systemic flaw: positive publication bias. Scientists rarely publish their failures.
Because the data landscape is heavily skewed toward successful outcomes, models learn the "paved center" of successful experiments but remain entirely blind to the boundaries of physical and chemical feasibility.
The Geometry of Over-Optimism
Why does a lack of negative data break machine learning models? The issue lies in how algorithms construct decision boundaries.
In a standard classification task, a model requires both positive examples (e.g., "this reaction worked") and negative examples (e.g., "this reaction failed") to draw a line between what is possible and what is not. When negative results are systematically omitted, the training data becomes what statisticians call positive-unlabeled. The model sees the successes, but the vast space of failures is left completely unmapped.
Standard Supervised Learning: Positive-Unlabeled (Scientific Literature):
+ + + | - - + + + | ? ?
+ + | - - + + | ? ?
+ + + | - - + + + | ? ?
^ ^
Clear Boundary Where does it end?
Without explicit negative labels to constrain it, a model's predicted boundary of feasibility expands unchecked. This mathematical distortion has severe real-world consequences across multiple scientific domains:
- Model Hallucinations: As Mroz et al. explain, the absence of negative data causes chemical AI models to become overly optimistic, frequently hallucinating unrealistic chemical properties or reaction outcomes that violate practical constraints.
- Inflated Confidence: Vasylenko et al. note that missing negatives in materials databases create a reporting bias that leads to highly inflated confidence scores, making models struggle to distinguish between high-performing and low-performing chemical systems.
- Blindness to Instability: Duan et al. emphasize that this systemic lack of negative data deprives models of the ability to accurately predict material instability or synthesizability, leading them to propose structures that look promising in theory but decompose instantly in practice.
- Reactivity Miscalculations: Kishimoto et al. and Schwaller et al. demonstrate that without access to failed experiments, reaction prediction models cannot learn the true boundaries of chemical reactivity, frequently recommending non-viable synthesis pathways.
This problem is not unique to chemistry. In biomedical research, Marcelo C. R. Melo points out that antibiotic discovery models become overly optimistic and fail to accurately predict molecular toxicity or inactivity because training sets lack molecules that failed to show activity. Similarly, Andreu-Sanz et al. show that positive publication bias in preclinical CAR-T cell therapy trials skews the perceived efficacy of new treatments, masking the boundaries of therapeutic success.
How LLMs Inherit and Amplify the Bias
The recent shift toward using Large Language Models (LLMs) to extract scientific knowledge has amplified this data scarcity problem. LLMs trained on millions of scientific papers do not just learn facts; they inherit the cultural biases of the publishing ecosystem.
Schilling-Wilhelmi et al. highlight that because scientific literature omits failed attempts, LLMs inherit a fundamental bias toward successful experiments. This creates what Subramanian et al. call an execution gap—the friction between digital design and physical realization. An LLM can easily generate millions of novel molecular structures, but because it lacks data on what fails, it cannot predict whether those molecules can actually be synthesized. As Miret et al. argue, this lack of negative results prevents models from identifying experimental constraints, leading autonomous discovery pipelines down unproductive research paths.
Furthermore, extracting what little negative data does exist in literature is incredibly difficult. Scientific narratives often mention failed reactions or model systems as context before presenting their primary, successful results. Bran et al. found that current text-extraction frameworks frequently fail to distinguish between these successful primary results and the auxiliary failures discussed in the text, creating a distorted map of chemical reactivity boundaries.
Strategies for Mapping the Failure Space
To build robust models, researchers are developing methods to systematically capture, generate, or algorithmically bypass the lack of negative data.
1. High-Throughput Experimentation and Closed-Loop Labs
The most direct solution is to change how scientific data is collected. In fields like protein crystallography, Newman et al. note that "failures" make up 99.8% of all experimental screening data. Instead of discarding these trials, researchers are leveraging autonomous self-driving labs—as outlined by Friederich et al.—to systematically document every failed attempt in machine-readable formats.
By pairing these automated labs with machine learning, models can be trained on balanced datasets. Zhong et al. demonstrate that incorporating expert-designed negative results from high-throughput experimentation (HTE) into Bayesian deep learning models allows them to accurately map the boundaries of chemical feasibility, vastly outperforming models trained purely on success-biased literature.
2. Explicit Failure Integration
When models are intentionally fed negative data, their predictive accuracy improves dramatically. For instance, Jiang et al. show that explicitly incorporating failed reaction data into deep chemical language processing models allows them to accurately judge reaction practicality, successfully separating reactions that appear theoretically sound from those that are practically unfeasible.
Similarly, Ignácz et al. generated an open-access dataset containing both successful and failed interfacial polymerizations. They proved that training models on datasets that explicitly include these negative results allows virtual screening tools to learn the true boundaries of a chemical design space rather than just memorizing the successful regions.
3. One-Class Classification
When negative data is completely unavailable and cannot be generated, researchers are changing the algorithmic framing of the problem. Instead of forcing a model to perform binary classification (success vs. failure) on biased data, they treat it as a one-class classification problem, where the model only learns the distribution of the positive data and flags anything outside that distribution as a potential failure.
To make this approach viable, Zhang et al. developed the Maximum Fraction Difference (MFD) method. MFD evaluates one-class classification models using only positive data, allowing researchers to accurately predict the synthesizability of complex materials like metal-organic frameworks (MOFs) despite the total absence of published negative results.
Conclusion
For machine learning to truly revolutionize scientific discovery, models must understand not just what works, but what fails. Whether through the deployment of self-driving labs that log every error, text-mining tools that capture auxiliary failures, or mathematical frameworks designed for one-class data, mapping the boundaries of failure is the essential next step in building AI systems that can safely and accurately navigate the physical world.