Heather Kulik

Heather J. Kulik is MIT’s Lammot du Pont Professor of Chemical Engineering, with a joint appointment in Chemistry. Her research combines quantum chemistry and machine learning to guide molecular and materials discovery. In a 2025 collaboration with Duke University, she and lead author Ilia Kevlishvili developed computational screening that identified a crosslinker producing a polymer about four times tougher than the standard-ferrocene comparison material.[2][3][9]

Kulik and Duke chemist Stephen Craig pursued mechanophores—molecules that change in response to mechanical force—as weak crosslinkers between polymer strands. Their work built on a 2023 finding by Craig, MIT’s Jeremiah Johnson, and colleagues: weaker crosslinks can increase a network’s resistance to tearing. A propagating crack preferentially follows the weaker bonds, forcing it to break more bonds than it would in a network with uniformly strong links.[9]

The team focused on ferrocenes, compounds with an iron atom between two carbon-containing rings. They began with about 5,000 already-synthesized structures in the Cambridge Structural Database; Kevlishvili explained that this reduced uncertainty about whether the mechanophores themselves could be made. After simulating about 400 compounds, they trained a neural network to predict the force needed to activate the remaining roughly 4,500, plus another 7,000 candidates with rearranged atoms. Craig credited Kulik and Kevlishvili with identifying the screening challenge and devising the approach.[9]

The screening indicated that bulky groups attached to both ferrocene rings could make force-induced breakup more likely. “This was something truly surprising,” Kulik told MIT News. Craig’s Duke laboratory incorporated a selected compound, m-TMS-Fc, into a polyacrylate network. The researchers reported about fourfold toughness compared with polymers using standard ferrocene crosslinkers; the research abstract specifies greater than fourfold enhancement in material tearing energy. Kulik was senior author, with collaborators including Kevlishvili, Jafer Vakil, David Kastner, Xiao Huang, and Craig.[9][10]

Kulik questions expectations of an AlphaFold-like breakthrough in materials, emphasizing sparse experimental data, approximate density functional theory (DFT) reference calculations, and the breadth of interactions introduced by different elements. Her assessment, quoted in Brandon Anderson and RJ Honicky’s March 24, 2026 Latent Space interview recap, is: “We have really good datasets for really boring chemistry.”[1]

Kulik tests chemical constraint-following by asking updated LLMs to design a ligand containing exactly twenty-two heavy atoms. At the time of recording, she reported repeated failures, according to the interview recap. In subsequent tests, the editors obtained correct kinase-ligand designs from Claude and ChatGPT, but their metal-organic framework (MOF) requests produced ligands with 21, 23, or 24 atoms.[1]

Kulik urges care when mining scientific literature: her team sometimes found disagreements between reported property values and the graphs in papers, according to the interview recap. She also advocates academic curiosity about questions outside heavily capitalized AI companies’ agendas, with researchers building shared resources such as organized data, high-throughput experimentation, and compute.[1]

Earlier collaborations provide context for these concerns: Kulik worked on predicting quantum-chemical properties, extracting experimental stability measurements from papers, and choosing among computational approximations before the ferrocene study.[6][7][8]

Kulik earned a BE in chemical engineering from Cooper Union in 2004 and a PhD in materials science and engineering from MIT in 2009. She completed postdoctoral training at Lawrence Livermore National Laboratory and Stanford University before joining MIT’s Chemical Engineering faculty in November 2013. MIT Chemistry announced her joint appointment in December 2022; her named professorship was announced in October 2024.[3][4]

In 2017 work with Jon Paul Janet, Kulik coauthored neural-network models for transition-metal-complex spin-state ordering, sensitivity to Hartree–Fock exchange, and spin-state-specific bond lengths. The models used empirical descriptors suited to inorganic chemistry without requiring precise three-dimensional structures. The authors reported spin-state splittings within 3 kcal/mol of DFT calculations for single-site chromium-through-nickel complexes at varying amounts of Hartree–Fock exchange, alongside tools for estimating prediction uncertainty.[6]

In 2021, Aditya Nandy, Chenru Duan, and Kulik used natural language processing and automated image analysis to extract MOF stability data from nearly 4,000 manuscripts. They obtained over 2,000 solvent-removal stability measures and 3,000 thermal degradation temperatures, then trained Gaussian-process and neural-network models using graph and pore-structure representations. Interpreting the models’ important features helped them identify strategies for improving the stability of MOFs containing 3d metals.[7]

A December 2022 paper by Chenru Duan, Aditya Nandy, Ralf Meyer, Naveen Arunachalam, and Kulik addressed the choice of density functional approximation. Their recommender combined electron-density fitting and delta-learning to select the approximation with the lowest expected error relative to costly coupled-cluster calculations for a particular system. For vertical spin-splitting energies of transition-metal complexes, the authors reported accuracy of about 2 kcal/mol, outperforming individual delta-learning models and the best conventional single-functional approach from a set of 48 approximations.[8]