fix: clean up phonons tab info

app.py

@@ -411,8 +411,34 @@ def predict_jdft2d(formula: str):
     if err: return f"❌ {err}", ""
     return f"### {pred:.1f} meV/atom", f"**{pred:.1f} meV/atom** exfoliation"
 
-def predict_phonons_placeholder(formula: str):
-    return "### ⚠️ Structure Required", "Phonon peak frequency (cm⁻¹) requires CIF file."
+PHONONS_INFO = """
+## 🎵 Phonon Peak Frequency
+
+The **TRIADS V6 Graph-TRM** achieves **41.91 ± 4.04 cm⁻¹ MAE** on Matbench phonons, using a gate-based halting Graph Neural Network that adaptively runs 4–16 message-passing cycles.
+
+### Architecture
+- **Gate-based halting**: 4–16 adaptive GNN cycles (halts when gate activations drop below threshold)
+- **Graph Attention TRM**: line-graph bond updates + joint self-attention + cross-attention
+- **Input**: Full crystal structure — atom positions, bond distances, angles (requires CIF/POSCAR)
+
+### Why no live demo?
+The phonons model requires a **pre-computed crystal graph** (atom positions, bond lengths, bond angles). 
+Composition-only featurization is insufficient for phonon prediction — structural details like bond stiffness 
+and crystal symmetry are essential.
+
+### Benchmark Results
+| Model | MAE (cm⁻¹) |
+|---|---|
+| **TRIADS V6 (ours)** | **41.91 ± 4.04** |
+| MEGNet | 28.76 |
+| ALIGNN | 29.34 |
+| MODNet | 45.39 |
+| CrabNet | 47.09 |
+| TRIADS V4 | 56.33 |
+
+> **Note**: MEGNet and ALIGNN use full DFT structural relaxation data.
+> TRIADS V6 achieves competitive performance with a simpler, more parameter-efficient Graph-TRM architecture (< 50K parameters).
+"""
 
 
 # ─────────────────────────────────────────────────────────────────
@@ -466,11 +492,7 @@ def build():
                 btn_j.click(predict_jdft2d, f_j, [out_j, ctx_j])
 
             with gr.Tab("🎵 Phonons"):
-                f_ph = gr.Textbox(label="Formula", value="Si")
-                btn_ph = gr.Button("Show info", variant="primary")
-                out_ph = gr.Markdown(elem_id="result_text")
-                ctx_ph = gr.Markdown()
-                btn_ph.click(predict_phonons_placeholder, f_ph, [out_ph, ctx_ph])
+                gr.Markdown(PHONONS_INFO)
 
     return demo